fix: litellm-health step 8 - run key/list on CT 116 host, not in container

- harness-litellm container has no curl/wget, so docker exec harness-litellm curl returns empty
- Fix: run curl on CT 116 host (ssh root@192.168.68.116 then curl)
- Add admin-call-failed label for empty/unparseable responses
- Verify: 10 keys found (host-side curl), NO-CURL confirmed in container

.gitea/workflows/pr-pipeline.yaml

@@ -43,17 +43,21 @@ jobs:
           echo "=== Prose Contract Frontmatter Validation ==="
           FAILED=0
           for f in $(find . -name "*.prose.md" -not -path "./.git/*" -not -path "./runs/*"); do
+            # NOTE: use herestrings, not `echo "$FM" | grep ...`. Under the runner's
+            # `-e -o pipefail`, `grep -q` exits on first match and can SIGPIPE the
+            # producer, making the pipeline report non-zero and raising a false
+            # "Missing name/description" whose file set varies run to run.
             FM=$(sed -n '/^---$/,/^---$/p' "$f" | sed '1d;$d')
             [ -z "$FM" ] && { echo "  ❌ $f: No YAML frontmatter"; FAILED=$((FAILED+1)); continue; }
 
-            KIND=$(echo "$FM" | grep '^kind:' | awk '{print $2}')
+            KIND=$(grep '^kind:' <<< "$FM" | awk '{print $2}')
             case "$KIND" in
               function|responsibility|gateway|pattern|test|template|architecture|enforcement) echo "  ✅ $f: kind=$KIND" ;;
               *) echo "  ❌ $f: Invalid kind='$KIND'"; FAILED=$((FAILED+1)) ;;
             esac
 
-            echo "$FM" | grep -q '^name:' || { echo "  ❌ $f: Missing name"; FAILED=$((FAILED+1)); }
-            echo "$FM" | grep -q '^description:' || { echo "  ❌ $f: Missing description"; FAILED=$((FAILED+1)); }
+            grep -q '^name:' <<< "$FM" || { echo "  ❌ $f: Missing name"; FAILED=$((FAILED+1)); }
+            grep -q '^description:' <<< "$FM" || { echo "  ❌ $f: Missing description"; FAILED=$((FAILED+1)); }
           done
           [ $FAILED -gt 0 ] && { echo "❌ FRONTMATTER FAILED ($FAILED error(s))"; exit 1; }
           echo "✅ Frontmatter validation passed"

AGENTS.md

@@ -55,7 +55,7 @@ Two incidents taught us this:
 ### Stage 3 — AI Review
 - Diff is sent to `syslog-auto` model via LiteLLM
 - Review checks against infrastructure-control ground truth:
-  - CT IDs match PVE cluster (100-117, no 122/123)
+  - CT IDs match the PVE cluster inventory in `infrastructure-control.prose.md` Appendix B (100-120 with gaps; no 122/123)
   - Grafana is direct LAN :3001, NOT behind nginx
   - Zulip is CT 117 on storepve (bridge IP .19)
   - Strix Halo :8080 is firewalled to .116 only

README.md

@@ -88,7 +88,7 @@ prose run memory-audit-maintenance memory_threshold=90 verify_configs=true
 prose run hermes-config-template agent_name=syslog-devops default_model=claude-sonnet-4
 
 # Configure an agent with a different auxiliary model
-prose run hermes-config-template agent_name=syslog-code default_model=qwen3.6-27B-code auxiliary_model=gpu-vision
+prose run hermes-config-template agent_name=syslog-code default_model=gpu-dense auxiliary_model=gpu-vision
 ```
 
 ### Option B: Manual Execution

audit-hermes-config.py

@@ -36,6 +36,59 @@ def warn(rule, message):
     WARNINGS.append(f"[{rule}] {message}")
 
 
+# Derivation rule: a model name is any scalar under a mapping key named `model` or
+# `model_name`, at any depth. The top-level `model:` SECTION is the exception where the model
+# name lives under `default`/`model`/`model_name` inside that section, so it is descended
+# specially. The only other exception is key `models` (litellm key-generation params carry a
+# list of model names). EXTEND THE ALLOWLIST for a new exception; do NOT add another field by
+# hand.
+MODEL_KEYS = ("model", "model_name")
+MODEL_SECTION_KEYS = ("default", "model", "model_name")
+MODEL_LIST_KEYS = ("models",)
+
+
+def _iter_model_values(node, path=""):
+    """Yield (path, value) for every model-name-bearing scalar in a config."""
+    if isinstance(node, dict):
+        for key, value in node.items():
+            child = f"{path}.{key}" if path else key
+            if key in MODEL_KEYS:
+                if isinstance(value, dict):
+                    for subkey in MODEL_SECTION_KEYS:
+                        subvalue = value.get(subkey)
+                        if isinstance(subvalue, str):
+                            yield (f"{child}.{subkey}", subvalue)
+                    for subkey, subvalue in value.items():
+                        if isinstance(subvalue, (dict, list)):
+                            yield from _iter_model_values(subvalue, f"{child}.{subkey}")
+                elif isinstance(value, list):
+                    yield from _iter_model_values(value, child)
+                else:
+                    yield (child, value)
+            elif key in MODEL_LIST_KEYS:
+                yield from _iter_model_list(value, child)
+            elif isinstance(value, (dict, list)):
+                yield from _iter_model_values(value, child)
+    elif isinstance(node, list):
+        for i, item in enumerate(node):
+            yield from _iter_model_values(item, f"{path}[{i}]")
+
+
+def _iter_model_list(node, path):
+    """Yield scalars under an allowlisted `models` key (list of names or list of dicts)."""
+    if isinstance(node, list):
+        for i, item in enumerate(node):
+            yield from _iter_model_list(item, f"{path}[{i}]")
+    elif isinstance(node, dict):
+        for key, value in node.items():
+            if key in MODEL_KEYS and isinstance(value, str):
+                yield (f"{path}.{key}", value)
+            elif isinstance(value, (dict, list)):
+                yield from _iter_model_list(value, f"{path}.{key}")
+    else:
+        yield (path, node)
+
+
 def audit(path):
     with open(path) as f:
         cfg = yaml.safe_load(f)
@@ -89,15 +142,16 @@ def audit(path):
     )
 
     # --- Rule 8: GPU Workload Distribution ---
+    # gpu-light (and gemma-4-12b) were retired 2026-09-12; the RTX 5070 stable alias is gpu-vision.
     check(
-        aux.get("vision", {}).get("model") == "gpu-light",
+        aux.get("vision", {}).get("model") == "gpu-vision",
         "Rule 8",
-        f"auxiliary.vision.model must be gpu-light (got {aux.get('vision', {}).get('model')!r}) — RTX 5070 stable alias",
+        f"auxiliary.vision.model must be gpu-vision (got {aux.get('vision', {}).get('model')!r}) — RTX 5070 stable alias",
     )
     check(
-        aux.get("web_extract", {}).get("model") == "gpu-light",
+        aux.get("web_extract", {}).get("model") == "gpu-vision",
         "Rule 8",
-        f"auxiliary.web_extract.model must be gpu-light (got {aux.get('web_extract', {}).get('model')!r}) — RTX 5070 stable alias",
+        f"auxiliary.web_extract.model must be gpu-vision (got {aux.get('web_extract', {}).get('model')!r}) — RTX 5070 stable alias",
     )
 
     # --- Rule 9: Compression Threshold ---
@@ -109,7 +163,7 @@ def audit(path):
     check(
         comp.get("max_context_window") == 131072,
         "Rule 9",
-        f"compression.max_context_window must be 131072 (got {comp.get('max_context_window')!r}) — matches 128K GPU capacity",
+        f"compression.max_context_window must be 131072 (got {comp.get('max_context_window')!r}) — syslog-auto pool floor (NVIDIA hosts 128K; Strix Halo 256K)",
     )
 
     # --- Rule 10: Default Model Must Be syslog-auto ---
@@ -178,21 +232,34 @@ def audit(path):
         f"custom_providers[0].base_url must end with /v1 (got {cp.get('base_url')!r})",
     )
 
-    # --- No raw model names (Rule 7/8 spirit) ---
-    raw_names = {"gemma-4-12b", "qwen3.6-27B-code", "qwen3.6-35B-udq4", "ornith-1.0-35b"}
-    for section_path, section_dict in [
-        ("model", model), ("compression", comp),
-        ("auxiliary.vision", aux.get("vision", {})),
-        ("auxiliary.web_extract", aux.get("web_extract", {})),
-        ("auxiliary.compression", aux.get("compression", {})),
-        ("delegation", deleg),
-    ]:
-        m = section_dict.get("model", "")
-        if m in raw_names:
+    # --- Retired/raw model names (Rule 7/8 spirit) ---
+    # The audit's job is to catch configs that are BROKEN, not to enforce a style preference.
+    # NON-RESOLVING names (removed 2026-09-12, verified 400/403 via live LiteLLM) must hard-FAIL:
+    #   gpu-light -> gpu-vision ; gemma-4-12b -> gpu-vision
+    #   crew-auto -> syslog-auto (its 64K cap is retired; no cap in force) ; ornith-1.0-35b -> strix-moe
+    # RESOLVING names (verified 200) are discouraged but working, so they only WARN:
+    #   qwen3.6-27B-code -> gpu-dense ; qwen3.6-35B-udq4 -> strix-moe
+    # Failing a working alias would reject valid configs - the exact defect this change fixes.
+    non_resolving = {
+        "gpu-light": "gpu-vision",
+        "gemma-4-12b": "gpu-vision",
+        "crew-auto": "syslog-auto (its 64K cap is retired; no cap in force)",
+        "ornith-1.0-35b": "strix-moe",
+        "qwen3.6-27B-code": "gpu-dense",
+        "qwen3.6-35B-udq4": "strix-moe",
+    }
+    raw_but_live = {}
+    for field_path, value in _iter_model_values(cfg):
+        if value in non_resolving:
+            check(
+                False,
+                "Rule 7/8",
+                f"{field_path} = {value!r} is retired and no longer resolves (2026-09-12) — use {non_resolving[value]}",
+            )
+        elif value in raw_but_live:
             warn(
                 "Rule 7/8",
-                f"{section_path}.model = {m!r} — raw model name, use stable alias instead "
-                f"(gpu-light, gpu-dense, strix-moe, syslog-auto)",
+                f"{field_path} = {value!r} is a raw-but-live model name — prefer the stable alias {raw_but_live[value]}",
             )
 
     # --- Report ---

disk-gc-threat-response.prose.md

@@ -1,11 +1,14 @@
 ---
 report_only_agents:
   - koby  # ⛔ KOBY IS NEVER REPAIRED (Rule 17, 2026-08-17) — detect + report, never fix on .129
+# ⛔ The guest/host-keyed report-only gate the GC executor MUST honour lives in the body
+# "Hard gate" YAML block below — that block is authoritative and is the only copy.
 kind: responsibility
 name: disk-gc-threat-response
 description: >
   Recurring disk health scan, garbage collection, and threat response
-  across 15 Proxmox CTs + 3 GPU bare-metal hosts. Triggered by incident
+  across 20 Proxmox guests (17 LXC + 3 QEMU VMs) + 3 GPU bare-metal hosts
+  (fleet verified against `pvesh get /cluster/resources` 2026-09-12). Triggered by incident
   2026-07-04 where CT 105 (kagentz) hit 87% disk (49G/59G) from
   Docker image bloat — 5 dangling images, 15 build cache layers.
   Recovered 35.67GB. Second incident 2026-07-09: amdpve (.15) Docker
@@ -25,7 +28,7 @@ logged within 5 minutes of discovery.
 
 ## Scope
 
-All 15 CTs via `pct-run` + 3 GPU bare-metal hosts via direct SSH.
+All 20 Proxmox guests (17 LXC via `pct-run` + 3 QEMU VMs via direct SSH) + 3 GPU bare-metal hosts via direct SSH.
 Docker hosts get special attention:
 
 | Host | CT | Disk Risk | GC Strategy |
@@ -99,35 +102,67 @@ and escalation trail.
 
 ## Execution
 
+### Hard gate: report-only guests (READ THIS BEFORE RUNNING GC)
+
+**CT 111 / hostname `tdunna` / 192.168.68.129 is DETECT-AND-REPORT-ONLY.** It belongs to Theo.
+The captain ruled 2026-08-17 and re-confirmed 2026-09-10 that Theo handles CT 111 himself.
+At **every** threat level — AMBER, RED, or CRITICAL — the executor must:
+
+- push the threat row and alert the owner, and
+- **never** call `gc-executor`, and **never** run any GC command against that guest: no
+  `apt-get clean/autoremove`, no `journalctl --vacuum-*`, no `find /var/log -delete`, no
+  `/tmp`/`/var/tmp` deletion, no snap removal, no `docker system prune`.
+
+This gate is keyed on **guest id / hostname / IP**, not on an agent name. The frontmatter
+`report_only_agents` marker (e.g. `koby`) names an AGENT while the scan unit is a GUEST, so an
+agent-name marker can silently miss the guest it lives on — it must never be the only gate.
+
+**The authoritative machine-readable exclusion list is the YAML block below.** The executor
+reads it at run time; `scripts/disk-gc-plan.py` turns a fleet scan into the action plan using it.
+Extend the list here, never by hand-maintaining a second copy. The Execution loop below MUST
+call that planner and MUST NOT reimplement the gate.
+
+```yaml
+# disk-gc report-only guests — authoritative. Keyed on guest/host, not agent.
+report_only_guests:
+  - guest: 111
+    hostname: tdunna
+    ip: 192.168.68.129
+    node: storepve
+    reason: "Theo's box — captain ruling 2026-08-17, re-confirmed 2026-09-10"
+```
+
+### Loop
+
 ```prose
 let fleet = call disk-scanner
   scope: all
 
-let threats = []
-for ct in fleet:
-  if ct.usage_pct >= 95:
-    push threats { ct: ct.id, level: "CRITICAL", pct: ct.usage_pct }
-  else if ct.usage_pct >= 85:
-    push threats { ct: ct.id, level: "RED", pct: ct.usage_pct }
-  else if ct.usage_pct >= 75:
-    push threats { ct: ct.id, level: "AMBER", pct: ct.usage_pct }
+-- The report-only gate is IMPLEMENTED IN scripts/disk-gc-plan.py and MUST NOT be
+-- reimplemented here. That planner reads the contract's `report_only_guests` YAML block
+-- and matches on guest id OR hostname OR IP, so the tested gate is the executed gate.
+let plan = call disk-gc-plan
+  fleet: fleet
 
--- sort by severity descending
-sort threats by pct desc
+for row in plan:
+  if row.action == "report-only":
+    -- Excluded guest: alert only. No gc-executor call is constructed for it, at any level.
+    call alerter
+      threat: row
+      result: { action: "report-only", reason: row.reason }
+  else:
+    let result = call gc-executor
+      ct: row.target
+      level: row.level
+      strategy: lookup-gc-strategy(row.target)
 
-for threat in threats:
-  let result = call gc-executor
-    ct: threat.ct
-    level: threat.level
-    strategy: lookup-gc-strategy(threat.ct)
-  
-  call alerter
-    threat: threat
-    result: result
+    call alerter
+      threat: row
+      result: result
 
 call summary-reporter
   fleet: fleet
-  threats: threats
+  plan: plan
 ```
 
 ## GC Strategies by Host Type
@@ -239,7 +274,9 @@ dangling images and orphaned build cache. No automated GC was in place.
 ## Incident Log: 2026-07-09 — amdpve docker bloat
 
 ### Discovery
-Scheduled fleet disk scan via `pct-run` across all 15 CTs + 3 GPU bare-metal hosts.
+Scheduled fleet disk scan across all 20 Proxmox guests (17 LXC via `pct-run`, 3 QEMU VMs via direct SSH) + 3 GPU bare-metal hosts.
+> **Report-only gate applies to this scan:** CT 111 (`tdunna`, 192.168.68.129) is alerted but never
+> garbage-collected at any level.
 amdpve (.15) flagged at 78% (AMBER threshold: 75%).
 
 ### Diagnosis
@@ -272,25 +309,35 @@ one-off GPU builds. No automated post-migration cleanup was in place.
 - Contract now scans GPU bare-metal hosts alongside CTs
 - Access via `pct-run` script for all CTs (no hardcoded IPs)
 
-## Access Matrix (documented 2026-07-09)
+## Access Matrix (verified against `pvesh get /cluster/resources` 2026-09-12)
 
-### CT Access (via pct-run)
-| CT | Name | Node | Status |
-|----|------|------|--------|
-| 100 | abiba | minipve | local |
-| 102 | adguard | minipve | ✅ reachable |
-| 104 | authentik | minipve | ✅ reachable |
-| 105 | kagentz | minipve | ✅ reachable |
-| 106 | ra-h-os | storepve | ✅ reachable |
-| 107 | pbs | storepve | ✅ reachable |
-| 108 | media | storepve | ✅ reachable |
-| 110 | gitea | minipve | ✅ reachable |
-| 111 | tdunna | amdpve | ✅ reachable |
-| 112 | tanko | amdpve | ✅ reachable |
-| 113 | baggy | amdpve | ✅ reachable |
-| 115 | scottdenya | amdpve | ✅ reachable |
-| 116 | syslog-api | minipve | ✅ reachable |
-| 117 | zulip | storepve | ✅ reachable |
+### Guest Access (via `pct-run` — CT id only, node resolved by `scripts/pct-run.sh`)
+| Guest | Name | Node | Type | Status |
+|------|------|------|------|--------|
+| 100 | abiba | minipve | lxc | ✅ reachable (probed via pct-run like any other guest; no local shortcut) |
+| 102 | adguard | minipve | lxc | ✅ reachable |
+| 104 | authentik | minipve | lxc | ✅ reachable |
+| 105 | kagentz | **amdpve** | lxc | ✅ reachable (was documented as minipve — corrected) |
+| 106 | ra-h-os | storepve | lxc | ✅ reachable |
+| 107 | pbs | storepve | lxc | ✅ reachable |
+| 108 | media | storepve | lxc | ✅ reachable |
+| 110 | gitea | minipve | lxc | ✅ reachable |
+| 111 | tdunna | **storepve** | lxc | ⛔ **REPORT-ONLY** (192.168.68.129, Theo's box — no GC at any level) |
+| 112 | tanko | amdpve | lxc | ✅ reachable |
+| 113 | baggy | amdpve | lxc | ✅ reachable |
+| 115 | scottdenya | amdpve | lxc | ✅ reachable |
+| 116 | syslog-api | minipve | lxc | ✅ reachable |
+| 117 | zulip | storepve | lxc | ✅ reachable |
+| 118 | jdownloader | storepve | lxc | ✅ reachable |
+| 119 | infisical-vault | minipve | lxc | ✅ reachable |
+| 120 | adguard2 | amdpve | lxc | ✅ reachable |
+
+### QEMU VMs (via direct SSH)
+| VM | Name | Node | IP | Status |
+|----|------|------|-----|--------|
+| 101 | llm-gpu (workload now bare metal .8) | acerpve | — | ✅ reachable |
+| 103 | ocu-llm (workload now bare metal .110) | ocupve | — | ✅ reachable |
+| 109 | docker-vm | storepve | 192.168.68.7 | ✅ reachable |
 
 ### GPU Bare Metal (via direct SSH)
 | Host | IP | GPU | Status |
@@ -299,9 +346,14 @@ one-off GPU builds. No automated post-migration cleanup was in place.
 | ocu-llm | 192.168.68.110 | RTX 5070 | ✅ reachable |
 | amdpve | 192.168.68.15 | Strix Halo | ✅ reachable |
 
-### KVM VM (via direct SSH)
-| Host | IP | Role | Status |
-|------|-----|------|--------|
-| docker-vm | 192.168.68.7 | 16 Docker containers, 4 stacks | ✅ reachable |
-
-> **Note:** CT 118 is now jdownloader (active on storepve). CT 119 (infisical-vault) added on minipve.\n> **Migrated:** CT 101 → .8, CT 103 → .110 (bare metal GPU).\n> **KVM VM:** CT 109 (docker-vm) is a KVM VM, not LXC — access via SSH .7.
+> **Fleet count:** 20 Proxmox guests (17 LXC + 3 QEMU VMs) + 3 GPU bare-metal hosts. Corrected
+> 2026-09-12: CT 105 → amdpve, CT 111 → storepve, and guests 118/119/120 were missing.
+>
+> **CT 100 probe gap (folded in):** CT 100 previously reported "unreachable (not reported)" every
+> run. Root cause is the same stale access layer: `pct-run` resolves the guest's node from its map,
+> and the map/contract must reflect `pvesh /cluster/resources`. Verified working from inside CT 100:
+> `scripts/pct-run.sh 100 "df -P / | tail -1"` → `23% /`. Probe CT 100 through `pct-run` like any
+> other guest — never through a local-only path, since the scanner itself runs inside CT 100 and a
+> container has no `pct` binary.
+>
+> **KVM VM:** CT 109 (docker-vm) is a QEMU VM, not LXC — access via SSH .7.

docs/probe-drift-round2-evidence.md

@@ -261,6 +261,10 @@ repaired.
    not exist` on .15. `agent-health-check.py` now carries the live-verified
    `storepve` mapping (the script is not the topology source of truth); the
    CRITICAL contract itself needs an authorized correction.
+   **✅ Resolved 2026-09-12:** the topology was corrected in its owner,
+   `infrastructure-control.prose.md` (CT 111 → storepve, CT 105 → amdpve), and
+   `scripts/pct-run.sh` now matches. This snapshot is left as observed; treat
+   those owner documents as authoritative.
 2. **Strix Halo `:8080` firewall claim is stale.** `prose-ai-review.sh`
    ground-truth rule #4 and `gpu-monitor.prose.md` say `:8080` is firewalled to
    `.116` only and `.24` cannot probe it. Live on .15:

gpu-fleet.prose.md

@@ -8,12 +8,12 @@ description: >
   UPDATED 2026-07-15: Stable role-based aliases introduced: strix-moe,
   gpu-dense, gpu-vision (gpu-light was superseded by gpu-vision on 2026-09-12).
   These never change — only the underlying model does.
-  Strix Halo: strix-moe → unsloth/Qwen3.6-35B-A3B-MTP (UD-Q4_K_M, 22GB).
+  Strix Halo: strix-moe → Carnice-Qwen3.6-MoE-35B-A3B-Q4_K_M.gguf (22GB, 256K ctx).
   RTX 5070: gpu-vision — IQ4_NL + MTP draft (~122 tok/s, 2x faster).
-  UPDATED 2026-07-17: Context reduced fleet-wide from 256K to 128K for stability.
-  Strix Halo model swapped to qwen3.6-35B-udq4 (22GB, strix-moe alias).
-  Instability observed near 100K at 256K (now all GPUs at 128K). 128K is the stable ceiling.
-  For larger context needs → fall back to external providers (deepseek).
+  UPDATED 2026-07-17: NVIDIA host context reduced from 256K to 128K for stability.
+  Strix Halo model: Carnice-Qwen3.6-MoE-35B-A3B-Q4_K_M.gguf (alias strix-moe, 256K context).
+  Strix Halo runs 256K (n_ctx 262144, --kv-unified); RTX 3090 and RTX 5070 remain at 128K.
+  For >128K on NVIDIA hosts → fall back to external providers (deepseek).
   VRAM headroom improved: RTX 3090 ~70%, RTX 5070 ~65%.
 agent: abiba
 triggers:
@@ -91,8 +91,8 @@ Single source of truth for models, aliases, rpm caps, weights and fallback chain
 CT 116 `/opt/inference-harness/litellm_config.yaml`. Do not duplicate those values in
 contracts — read them there.
 
-**Backward compatibility**: Old model-specific names (qwen3.6-27B-code, qwen3.5-9b-it) still work
-but are deprecated for agent configs. Only the stable aliases survive model swaps. `gemma-4-12b`
+**No backward compatibility**: Old model-specific names (qwen3.6-27B-code, qwen3.5-9b-it) are retired as of 2026-09-12
+and no longer resolve; do not use them in agent configs. Only the stable aliases survive model swaps. `gemma-4-12b`
 is retired and returns 400 `Invalid model name`.
 
 ## Routing Configuration (LiteLLM — July 2026)
@@ -190,7 +190,7 @@ If no SSH access, send Zulip DM via abiba-bot with vault update instructions.
 | `/root/scripts/gpu-saturation-watchdog.py` | pi (.24) | Auto-restart stuck llama-server |
 | `/root/dashboard/gpu-fleet.html` | pi (.24) | Live HTML dashboard |
 | `/etc/systemd/system/llama-server.service` | .8, .110 | llama-server daemons (Nvidia GPUs) |
-| `/etc/systemd/system/strix-server.service` | .15 (amdpve) | llama-server daemon (Vulkan, Strix Halo) running unsloth/Qwen3.6-35B-A3B-MTP-GGUF. Note: `llama-server.service` and `llama-server@.service` are **masked** on .15 to prevent port 8080 collisions. |
+| `/etc/systemd/system/strix-server.service` | .15 (amdpve) | llama-server daemon (Vulkan, Strix Halo) running Carnice-Qwen3.6-MoE-35B-A3B-Q4_K_M.gguf (256K ctx). Note: `llama-server.service` and `llama-server@.service` are **masked** on .15 to prevent port 8080 collisions. |
 
 ## Prometheus & Grafana
 
@@ -207,7 +207,7 @@ If no SSH access, send Zulip DM via abiba-bot with vault update instructions.
 - **Router startup race**: Compose router.py doesn't call load_roster(). Reload thread sleeps 30s first.
   Fix: trigger roster reload via SSH after restart, or rebuild image with startup load_roster().
 - **LiteLLM /metrics**: Requires auth. Prometheus uses `/health/liveliness` as workaround.
-- **Strix Halo GPU**: Vulkan is the working backend (ROCm/HIP path abandoned — HSA runtime blocked on Debian 13). Build at `/root/llama.cpp/build-vk/`, commit `4fc4ec5` (2026-07-01), ggml 0.15.3 shared-lib arch. Mesa RADV 25.0.7, KHR_coopmat fast path active. ~70 tok/s gen, 532 tok/s prompt. Service: `strix-server.service` on port 8080, model: `qwen3.6-35B-udq4`, alias `strix-moe`, 128K context, flash-attn + q4 KV, multimodal (mmproj loaded).
+- **Strix Halo GPU**: Vulkan is the working backend (ROCm/HIP path abandoned — HSA runtime blocked on Debian 13). Build at `/root/llama.cpp/build-vk/`, commit `4fc4ec5` (2026-07-01), ggml 0.15.3 shared-lib arch. Mesa RADV 25.0.7, KHR_coopmat fast path active. ~70 tok/s gen, 532 tok/s prompt. Service: `strix-server.service` on port 8080, model: `Carnice-Qwen3.6-MoE-35B-A3B-Q4_K_M.gguf`, alias `strix-moe`, 256K context (n_ctx 262144), --parallel 2 --kv-unified, flash-attn + q4 KV, multimodal (mmproj loaded).
 - **Port conflict detection (2026-07-05)**: All 3 GPU wrappers now detect ghost processes squatting port 8080 before starting. `.8` and `.110` use inline pre-start check in `llama-wrapper.sh`; `.15` uses `/usr/local/bin/port-cleanup.sh` ExecStartPre. Replaces the blanket `pkill -9 -x llama-server` on .15 which would kill ALL llama-server instances regardless of port. Ghost detection was the root cause of .8 crash-looping for 27+ restarts (stale pid 25836 squatting 8080 after OOM kill).
 - **Strix Halo thermal safeguard (2026-07-02)**: `strix-server.service` has `-n 8192` (hard generation cap per request). Without it, `--predict` defaults to -1 (infinity) — a runaway request from .123 (old Mumuni CT114 — now inside Abiba CT100 at .24) decoded 39,868 tokens over 24 min, pushing Tctl to 98°C (crit 89.8°C) and throttling 70→29 t/s. The cap bounds worst-case generation to ~5 min. Do NOT remove `-n` without a replacement ceiling. Sustained load hits ~84°C even at 92s; the APU is fanless/low-flow. Clients MUST also set `max_tokens`.
 - **Port 8080 firewall**: amdpve iptables restricts 8080 to 192.168.68.116 (LiteLLM/router host) only. All inbound connections are from .116 (LiteLLM proxied via nginx). Localhost curls hang (SYN dropped). Always test from .116.
@@ -216,17 +216,17 @@ If no SSH access, send Zulip DM via abiba-bot with vault update instructions.
 - **Alert migration**: All alerts now go to `#agent-hub` topics (`alerts-gpu`, `alerts-pm2`, `alerts-infra`) instead of DMs. Cross-agent visibility enabled.
 - **tok/s benchmarks**: Measured every 5 min via LiteLLM proxy. Baselines tracked with 30%/50% degradation thresholds.
 - **NetBird 502**: Tanko routes through NetBird for litellm.sysloggh.net. Use direct IP if NetBird down.
-- **Alias-retirement follow-up (2026-09-12)**: the agent-facing templates (`hermes-config-template.prose.md`, `hermes-agent-baseline.prose.md`), `litellm-api-keys.prose.md`, and the executable `audit-hermes-config.py` still reference the retired `gpu-light`/`gemma-4-12b` names, and `audit-hermes-config.py` Rule 8 currently fails a config whose vision/web_extract model is `gpu-vision`. These are tracked separately and are intentionally NOT updated in this change.
+- **Alias-retirement sweep (2026-09-12)**: `gemma-4-12b`, `gpu-light` and `crew-auto` are retired and replaced by `gpu-vision` / no cap respectively. The agent-facing templates (`hermes-config-template.prose.md`, `hermes-agent-baseline.prose.md`), `litellm-api-keys.prose.md`, `gpu-self-heal.prose.md`, `hermes-key-enforcement.prose.md`, `inference-optimization.prose.md`, `litellm-client-timeouts.prose.md` and the executable `audit-hermes-config.py` were all updated to the live canonical alias in the same change. **koby's config on .129 still names `gpu-light` (and `gemma-4-E4B`); .129 is report-only, so that is recorded for its owner and NOT edited here.**
 
 ## GPU Inference Benchmarks (Current)
 
 | GPU | Model | Gen tok/s | Prompt tok/s | Baseline | Context |
 |-----|-------|-----------|--------------|----------|---------|
 | RTX 3090 (.8) | Qwen3.8-27B-Uncensored-Q4_K_M | **TBD** | — | — | **128K** |
-| Strix Halo (.15) | qwen3.6-35B-udq4 | **65** | 140 | — | **128K** |
+| Strix Halo (.15) | Carnice-Qwen3.6-MoE-35B-A3B-Q4_K_M.gguf (strix-moe) | **65** | 140 | — | **256K** |
 
-Benchmarks from 2026-07-17. Strix Halo model: qwen3.6-35B-udq4. RTX 5070 MTP provides 2.7x speedup over pre-upgrade 70 tok/s.
-All 3 GPUs now at 128K context (2026-07-17, reduced from 256K for stability).
+Benchmarks from 2026-07-17. Strix Halo model: Carnice-Qwen3.6-MoE-35B-A3B-Q4_K_M.gguf (alias strix-moe), n_ctx 262144, --parallel 2 --kv-unified. RTX 5070 MTP provides 2.7x speedup over pre-upgrade 70 tok/s.
+GPU contexts: RTX 3090 (.8) and RTX 5070 (.110) at 128K; Strix Halo (.15) at 256K (2026-09-12).
 
 Benchmarks run through LiteLLM proxy (192.168.68.116:4000) every 5 minutes.
 Degradation alerts fire at 30% (warning) and 50% (critical) below baseline.
@@ -239,7 +239,7 @@ History stored at `/root/data/toks-history.json` with 7-day rolling window.
 ### Stable Aliases — CRITICAL
 
 All agent configs MUST use stable role-based aliases, never model-specific names:
-- `compression.model: strix-moe`
+- `compression.model: syslog-auto`
 - `auxiliary.vision.model: gpu-vision`
 - `delegation.model: gpu-dense`
 - `auxiliary.web_extract.model: gpu-vision`
@@ -247,27 +247,27 @@ All agent configs MUST use stable role-based aliases, never model-specific names
 When the underlying model is swapped, only the LiteLLM config changes — agent configs are untouched.
 
 ### Context Windows
-- RTX 3090: **128K** (reduced from 256K 2026-07-17) | RTX 5070: **128K** (reduced from 256K) | Strix Halo: **128K**
-- **All agents**: 128K ceiling — stable margin. For >128K workloads, use external providers (deepseek)
-- Compression threshold 0.60: fires at ~77K (~51K headroom before 128K ceiling)
+- RTX 3090: **128K** (reduced from 256K 2026-07-17) | RTX 5070: **128K** (reduced from 256K) | Strix Halo: **256K** (2026-09-12)
+- **Agents via `syslog-auto`**: 128K ceiling — the pool's safe floor (NVIDIA hosts are 128K). For >128K workloads, use external providers (deepseek)
+- Compression threshold 0.65 (audit Rule 9): fires at ~85K (~43K headroom before 128K ceiling)
 - **Pi agents (Abiba)**: `compaction.reserveTokens: 52739` (≈60% of 128K)
-- Mumuni compression model alias: `strix-moe`
+- Mumuni compression model alias: `syslog-auto`
 
 ### Mumuni Agent Profile
 
-Mumuni (kagentz CT105, 192.168.68.14 — migrated from CT100 2026-08-29) is the primary business assistant. This profile is the reference for all agent configs:
+Mumuni (kagentz CT105, 192.168.68.14 — migrated from CT100 2026-08-29) is the primary business assistant. This profile is the reference for all agent configs. The compression values below are the current required values per template Rules 7/9 and `audit-hermes-config.py`; whether Mumuni's LIVE config currently complies is a separate operational question.
 
 | Setting | Value | Notes |
 |---------|-------|-------|
 | `model.default` | `syslog-auto` | Balanced default (pool router) |
 | `model.provider` | `custom:litellm` | LiteLLM on CT116 |
-| `compression.model` | `strix-moe` | Stable alias — survives model swaps |
-| `aux.compression.model` | `strix-moe` | Compression auxiliary model |
+| `compression.model` | `syslog-auto` | Rule 7: auto-routing, prevents Strix Halo overload |
+| `aux.compression.model` | `syslog-auto` | Must match `compression.model` (Rule 7) |
 | `aux.vision.model` | `gpu-vision` | Vision tasks (RTX 5070) |
 | `aux.web_extract.model` | `gpu-vision` | Web extraction |
 | `delegation.model` | `gpu-dense` | Sub-agent reasoning (RTX 3090) |
-| `context.max_context_window` | 131072 (128K) | Reduced from 256K 2026-07-17 — stable 128K ceiling |
-| `compression.threshold` | 0.60 | Triggers at ~77K (~60% of 128K) — optimized for 128K context |
+| `context.max_context_window` | 131072 (128K) | Conservative `syslog-auto` pool floor (NVIDIA hosts 128K; Strix Halo 256K) |
+| `compression.threshold` | 0.65 | Rule 9: triggers at ~85K for a 128K window |
 | `compression.target_ratio` | 0.3 | Compresses to ~38K |
 | `compression.protect_last_n` | 40 | Preserves last 40 messages |
 | `memory.memory_char_limit` | 800 | Brief memory entries |

gpu-monitor.prose.md

@@ -27,8 +27,8 @@ agent: abiba
 ┌──────┐ ┌──────┐ ┌────────┐
 │.8:8080│ │.110  │ │.116:80 │
 │RTX3090│ │:8080 │ │nginx   │
-│gemma  │ │RTX5070│ │router  │
-└──────┘ │qwen27B│ │LiteLLM │
+│qwen   │ │RTX5070│ │router  │
+└──────┘ │vision │ │LiteLLM │
          └──────┘ │dashboard│
                   └────────┘
 ```

gpu-self-heal.prose.md

@@ -12,7 +12,7 @@ description: >
   Router (port 9000) DECOMMISSIONED 2026-09-11; references replaced with direct GPU routing.
   Benchmark baselines refreshed to live values.
   Prometheus exporters removed — not deployed; fall back to direct sidecar probes.
-  Stable role-based aliases (strix-moe, gpu-dense, gpu-light) from gpu-fleet.
+  Stable role-based aliases (strix-moe, gpu-dense, gpu-vision) from gpu-fleet.
 agent: abiba
 depends_on:
   - gpu-monitor.prose.md (live data source on .24:9100)
@@ -48,12 +48,12 @@ depends_on:
 
 | Alias | GPU | Host | Model | VRAM | Ctx | tok/s | Role |
 |-------|-----|------|-------|------|-----|-------|------|
-| `strix-moe` | Strix Halo 64GB | ct15 (.15:8080) | qwen3.6-35B-udq4 | ~10/64GB (16%) | 128K | 62.9 | Compression, summarization, long docs |
+| `strix-moe` | Strix Halo 64GB | ct15 (.15:8080) | Carnice-Qwen3.6-MoE-35B-A3B-Q4_K_M.gguf | ~10/64GB (16%) | 256K | 62.9 | Compression, summarization, long docs |
 
 Key notes:
 - All models use direct GPU routing via LiteLLM (`api_key: not-needed`). Router (port 9000) was decommissioned 2026-09-11 and is NOT in the inference path.
-- Stable aliases (gpu-dense, gpu-light, strix-moe) from gpu-fleet are the canonical names for agent configs. Model-specific names still work but are deprecated.
-- RTX 5070 tok/s is ~145 for Qwen3.5-9B — gpu-light is the fastest endpoint. Route vision/web/light work there first. NOTE: Qwen3.5-9B is multimodal (image+text), NOT text-only like gemma-4-12b was.
+- Stable aliases (gpu-dense, gpu-vision, strix-moe) from gpu-fleet are the canonical names for agent configs — use these, not model-specific names. The retired names `gpu-light` and `gemma-4-12b` were superseded by `gpu-vision` on 2026-09-12 and no longer resolve (400 `Invalid model name`).
+- The RTX 5070 is the fastest endpoint per token — `gpu-vision` is its canonical alias. Route vision/web/light work there first. The RTX 5070 model is multimodal (image+text).
 - Strix Halo is 62.9 tok/s (89% of 70.5 baseline) — below optimal but stable. Check for competing workloads.
 - RTX 3090 VRAM at 88% — within role-appropriate range (role = heavy reasoning, needs the headroom).
 - RTX 5070 VRAM at 83% — role-appropriate for vision/web (smaller batch sizes).
@@ -64,7 +64,7 @@ Key notes:
 - **Detect**: Any GPU temp >85°C sustained for 2+ consecutive polls
 - **Fix**: 
   1. Reduce inference concurrency on that GPU (load-side cooling only — NO fan control)
-  2. Redirect new requests to cooler GPUs via LiteLLM fallback chains (gemma → qwen, qwen → gemma)
+  2. Redirect new requests to cooler GPUs via LiteLLM fallback chains (gpu-vision → gpu-dense, gpu-dense → gpu-vision)
   3. If all GPUs hot, alert about cooling infrastructure
 - **Verify**: Temp drops below 80°C within 5 minutes
 - **Escalate after**: 3 verification failures → Zulip alert
@@ -142,10 +142,10 @@ Key notes:
 - **Escalate**: If Tier 2 triggers and temp still rising after 5 min → possible hardware failure
 
 ### Rule 9: Context Window Optimization
-- **Detect**: Benchmark tok/s vs baseline for each GPU at current context (all 128K)
+- **Detect**: Benchmark tok/s vs baseline for each GPU at its current context
   - RTX 3090 (128K ctx, ThinkingCap): baseline 74.8 tok/s — currently at 74.9 (100%)
   - RTX 5070 (128K ctx, HauhauCS QAT): baseline 165.2 tok/s — currently at 169.6 (103%)
-  - Strix Halo (128K ctx, qwen3.6-35B-udq4): baseline 70.5 tok/s — currently at 62.9 (89%)
+  - Strix Halo (256K ctx, Carnice-Qwen3.6-MoE-35B-A3B-Q4_K_M.gguf): baseline 70.5 tok/s — currently at 62.9 (89%)
 - **Fix**:
   - If tok/s > baseline → context has headroom, consider increasing
   - If tok/s < 90% baseline → reduce context by 25% and retest
@@ -157,13 +157,14 @@ Key notes:
 ### Rule 10: Workload Distribution Optimization (updated 2026-07-18)
 - **Detect**: GPU roles misaligned with hardware capabilities
 - **Target distribution**:
-  - RTX 3090 (gpu-dense, 24GB, 74.9 tok/s) → Heavy reasoning, code gen, long conversations (slowest per-token but largest context capacity). Weight: 0.55 (LiteLLM).
-  - RTX 5070 (gpu-light, 12GB, ~145 tok/s) → Vision (image+text), web search, lightweight tasks. Weight: 0.15 (LiteLLM).
-  - Strix Halo (strix-moe, 64GB, 62.9 tok/s) → Context compression, summarization, long docs (MoE model). Weight: 0.30 (LiteLLM).
+  - RTX 3090 (gpu-dense, 24GB, 74.9 tok/s) → Heavy reasoning, code gen, long conversations (slowest per-token but largest context capacity).
+  - RTX 5070 (gpu-vision, 12GB, ~145 tok/s) → Vision (image+text), web search, lightweight tasks.
+  - Strix Halo (strix-moe, 64GB, 62.9 tok/s) → Context compression, summarization, long docs (MoE model).
 - **Note**: RTX 5070 is the fastest endpoint per token. Route high-volume, low-complexity work there first.
+- **Weights are not restated here** — the live `syslog-auto` pool weights and rpm caps live in CT 116 `/opt/inference-harness/litellm_config.yaml`, the single source of truth.
 - **Fix**:
   - Alert if any GPU is handling workload outside its designated role
-  - Recommend agent alias updates to match workload to GPU role (use stable aliases: gpu-dense, gpu-light, strix-moe)
+  - Recommend agent alias updates to match workload to GPU role (use stable aliases: gpu-dense, gpu-vision, strix-moe)
   - Track per-GPU request distribution via LiteLLM spend logs
 - **Verify**: Each GPU's request pattern matches its designated role within 24h
 - **Escalate**: If role mismatch persists >48h → agent alias audit needed
@@ -311,6 +312,6 @@ Pushed to `SyslogSolution/health-logs/gpu/{run_id}.json` — versioned, searchab
   If monitor response > 1MB, log a warning and skip the cycle rather than crashing.
 
 ### L6: Stable Aliases Replace Model Names
-- gpu-fleet introduced stable aliases (strix-moe, gpu-dense, gpu-light) on 2026-07-15.
+- gpu-fleet introduced stable aliases (strix-moe, gpu-dense, gpu-light) on 2026-07-15; `gpu-light` was superseded by `gpu-vision` on 2026-09-12.
 - Self-heal must use aliases for reporting and alerting, not model-specific names.
 - **Rule**: All alert messages and KG nodes use the stable alias as the GPU identifier.

hermes-agent-baseline.prose.md

@@ -5,7 +5,7 @@ version: 1.0.0
 description: >
   Canonical known-good baseline for all Syslog Hermes agents. Captures the exact
   configuration state, keys, workarounds, and audit procedure. When an agent's
-  configuration goes sideways, restore from this baseline. Last verified 2026-07-16. All GPUs 128K context (reduced from 256K for stability Jul 2026) (RTX 3090 .8, RTX 5070 .110, Strix Halo .15). Parallel 1 fleet-wide (Strix Halo handles compression solo).
+  configuration goes sideways, restore from this baseline. Last verified 2026-07-16. RTX 3090/5070 at 128K (reduced from 256K for stability Jul 2026); Strix Halo at 256K (2026-09-12) (RTX 3090 .8, RTX 5070 .110, Strix Halo .15). Parallel 1 fleet-wide (Strix Halo handles compression solo).
 author: Abiba (pi agent)
 ---
 
@@ -24,11 +24,12 @@ done
 
 | Agent | CT | Node | IP | LiteLLM Alias | Key Source | Platform |
 |-------|-----|------|-----|---------------|------------|----------|
-| Koby | 111 | amdpve | .129 | `koby` | Infisical vault | **Hermes** |
+| Koby | 111 | storepve | .129 | `koby` | Infisical vault | **Hermes** |
 | Koonimo | 113 | amdpve | .114 | `koonimo` | Infisical vault | Hermes |
 | Shumba | — | 192.168.68.119 | N/A | N/A (DeepSeek) | Hermes (RETIRED — CT119 now Infisical vault) |
 
 > **Note**: CT hostnames (tdunna→CT111, baggy→CT113) differ from agent identities (koby, koonimo).
+> CT 111 (tdunna, 192.168.68.129, storepve) is report-only — Theo's box; alert only, never garbage-collect.
 
 Access: `pct-run <CT_ID> <command>` — no IPs needed. GPU hosts (.8, .110, .15) use SSH.
 Keys are stored in Infisical vault (project=agents, env=production) and injected at
@@ -80,7 +81,7 @@ custom_providers:
 auxiliary:
   vision:
     provider: harness
-    model: gemma-4-12b              # or syslog-auto
+    model: gpu-vision              # RTX 5070 stable alias (Rule 8; do not use syslog-auto for aux)
     base_url: http://192.168.68.116/litellm/v1
     api_key_env: LITELLM_API_KEY
     api_key: <value from: infisical secrets get LITELLM_API_KEY --project=agents --env=production>   # ← MANDATORY workaround
@@ -95,7 +96,7 @@ auxiliary:
     threshold: 0.65
     target_ratio: 0.3
     provider: harness
-    model: syslog-auto               # or gemma-4-12b
+    model: syslog-auto               # Rule 7: compression must be syslog-auto
     base_url: http://192.168.68.116/litellm/v1
     api_key_env: LITELLM_API_KEY
     api_key: <value from: infisical secrets get LITELLM_API_KEY --project=agents --env=production>   # ← MANDATORY workaround
@@ -185,7 +186,9 @@ Abiba (CT100) runs pi via PM2 with the Zulip extension.
 Config files: `~/.pi/agent/models.json`, `~/.pi/agent/settings.json`.
 
 **models.json** — Must only list models authorized for the agent's LiteLLM key.
-Key is injected via `infisical run --` wrapper at PM2 startup:
+`/v1/models` is key-scoped and the live registry is CT 116
+`/opt/inference-harness/litellm_config.yaml`; treat the list below as a snapshot and re-read
+the registry before applying. Key is injected via `infisical run --` wrapper at PM2 startup:
 ```json
 {
   "providers": {
@@ -197,9 +200,7 @@ Key is injected via `infisical run --` wrapper at PM2 startup:
         { "id": "syslog-auto" },
         { "id": "strix-moe" },
         { "id": "gpu-dense" },
-        { "id": "gpu-light" },
-        { "id": "qwen3.6-27B-code" },
-        { "id": "gemma-4-12b" }
+        { "id": "gpu-vision" }
       ]
     }
   }

hermes-config-template.prose.md

@@ -92,17 +92,17 @@ work immediately after restart.
 ```yaml
 # ─── Model Selection ───
 model:
-  default: <agent_model>          # e.g., strix-moe, qwen3.6-27B-code, syslog-auto
+  default: <agent_model>          # e.g., strix-moe, gpu-dense, syslog-auto
   provider: harness
   base_url: http://192.168.68.116/litellm/v1   # Rule 5 (2026-08-09): canonical authenticated path; /v1 also OK
   api_key_env: LITELLM_API_KEY    # Injected via infisical run -- wrapper
   max_tokens: 4096                # ⚠️ CRITICAL: Prevents unbounded generation
-  context_length: 131072          # For syslog-auto (all GPUs at 128K for stability).
+  context_length: 131072          # Conservative floor for syslog-auto (NVIDIA hosts 128K; Strix Halo 256K).
                                   # ⚠️ MANDATORY: Hermes probes unknown models from 256K
                                   # and falls back to 256K when /v1/models lacks a context
                                   # field (llama-server does). Without this override, agents
                                   # silently run syslog-auto at 256K (verified 2026-08-09).
-                                  # Set 65536 if using gemma-4-12b directly (tight VRAM).
+                                  # Set 65536 if pinning a single model directly (tight VRAM).
 
 fallback_providers:
   provider: deepseek
@@ -133,7 +133,7 @@ compression:
   enabled: true
   model: syslog-auto               # ⚠️ Must match auxiliary.compression.model. Stable alias (gpu-fleet § Stable Role-Based Aliases). NOT ornith-1.0-35b (LiteLLM does not serve that name).
   provider: harness
-  max_context_window: 131072      # MUST match actual GPU capacity. All 3 GPUs are 128K (Jul 17).
+  max_context_window: 131072      # MUST stay at the syslog-auto pool floor: NVIDIA hosts are 128K, Strix Halo 256K (2026-09-12).
   threshold: 0.65                 # Fires at ~170K for 262K window, ~85K for 128K
   target_ratio: 0.30
   protect_last_n: 40
@@ -143,25 +143,25 @@ compression:
 
 # ─── Auxiliary Tasks (CONSISTENCY RULE) ───
 # All auxiliary services MUST use identical model, base_url, and api_key_env:
-#   model: gpu-light                 # stable alias (NOT raw "gemma-4-12b")
+#   model: gpu-vision                # stable alias (NOT a raw model name)
 #   base_url: http://192.168.68.116/litellm/v1   # Rule 5 (2026-08-09): canonical authenticated; /v1 also OK
 #   api_key_env: LITELLM_API_KEY
 # Do NOT use syslog-auto for auxiliary tasks — it routes to the primary GPU.
-# gpu-light = RTX 5070 (12B), freeing the Strix Halo for agent reasoning.
+# gpu-vision = RTX 5070 (12B), freeing the Strix Halo for agent reasoning.
 # Heavy aux (delegation, x_search) use gpu-dense (RTX 3090) instead.
-# NEVER use raw model names (gemma-4-12b, qwen3.6-27B-code, qwen3.6-35B-udq4)
-# in agent configs — use the stable aliases so model swaps don't break agents.
+# NEVER use retired model names (qwen3.6-27B-code, qwen3.6-35B-udq4; gemma-4-12b is retired
+# and no longer resolves) in agent configs — use the stable aliases so model swaps don't break agents.
 auxiliary:
   vision:
     provider: harness
-    model: gpu-light            # stable alias for RTX 5070 (was raw gemma-4-12b)
+    model: gpu-vision           # stable alias for RTX 5070
     base_url: http://192.168.68.116/litellm/v1   # Rule 5 (2026-08-09): canonical authenticated; /v1 also OK
     api_key_env: LITELLM_API_KEY
     timeout: 60
     download_timeout: 30
   web_extract:
     provider: harness
-    model: gpu-light            # stable alias for RTX 5070
+    model: gpu-vision           # stable alias for RTX 5070
     base_url: http://192.168.68.116/litellm/v1   # Rule 5 (2026-08-09): canonical authenticated; /v1 also OK
     api_key_env: LITELLM_API_KEY
     timeout: 30
@@ -173,10 +173,10 @@ auxiliary:
     timeout: 300                 # gpu-fleet: 300s for large-history summarization (was 60)
 
 # ─── Delegation / Heavy Aux (use gpu-dense = RTX 3090) ───
-# delegation.model and x_search.model use gpu-dense (NOT raw qwen3.6-27B-code).
+# delegation.model and x_search.model use gpu-dense (NOT retired raw name).
 
 delegation:
-  model: gpu-dense              # stable alias for RTX 3090 (was raw qwen3.6-27B-code)
+  model: gpu-dense              # stable alias for RTX 3090
   provider: harness
   base_url: http://192.168.68.116/litellm/v1   # Rule 5 (2026-08-09): canonical authenticated; /v1 also OK
   api_key_env: LITELLM_API_KEY
@@ -252,10 +252,11 @@ The following MUST be identical across ALL profiles:
 - For agents needing longer outputs: raise to 8192, but never omit
 
 ### Rule 7: Auxiliary Model Consistency (UPDATED 2026-07-16)
-- Vision and web_extract use `gemma-4-12b` (RTX 5070 — 12GB, vision-optimized)
-- Compression uses `strix-moe` (stable alias for Strix Halo — 64GB, 128K ctx, compression-optimized)
-- **`strix-moe` is the only valid compression model name** — LiteLLM does NOT serve `ornith-1.0-35b`
-  (it serves `strix-moe`, `qwen3.6-35B-udq4`, `gpu-dense`, `gpu-light`, `syslog-auto`, `gemma-4-12b`, `qwen3.6-27B-code`). Old configs with `ornith-1.0-35b` cause 403/model-not-found on compression calls.
+- Vision and web_extract use `gpu-vision` (RTX 5070 — 12GB, vision-optimized)
+- Compression uses `syslog-auto` — the Strix Halo weighted pool (64GB, 256K ctx, compression-optimized); do NOT pin `compression.model` to `strix-moe` (audit Rule 7 rejects it)
+- **`ornith-1.0-35b` is NOT a valid compression model name** — LiteLLM does not serve it
+  (do not restate the served model list here — CT 116 `/opt/inference-harness/litellm_config.yaml`
+  is the single source of truth for models, aliases, weights and fallbacks). Old configs with `ornith-1.0-35b` cause 403/model-not-found on compression calls.
 - **OPERATIONAL DECISION (2026-07-23): Use `syslog-auto` for compression across all agents.**
   The `syslog-auto` alias routes to the Strix Halo, but uses the weighted pool instead of pinning
   to `strix-moe` directly. This prevents sustained Strix Halo thermal load because the pool can
@@ -264,41 +265,41 @@ The following MUST be identical across ALL profiles:
 - All auxiliary services MUST use identical routing:
   - `base_url: http://192.168.68.116/litellm/v1`  (Rule 5, 2026-08-09: canonical authenticated; `/v1` also OK)
   - `api_key_env: LITELLM_API_KEY`
-- **Do NOT use `syslog-auto`** for auxiliary tasks — it routes unpredictably
+- **Do NOT use `syslog-auto` for `vision`/`web_extract`** — it routes unpredictably; compression is the deliberate exception (see the OPERATIONAL DECISION above)
 - **Compression on Strix Halo**: The strix-moe alias routes to Strix Halo
-  (64GB UMA, 128K context) — the designated compression GPU. This frees the
+  (64GB UMA, 256K context) — the designated compression GPU. This frees the
   RTX 5070 for vision and web search, and the RTX 3090 for heavy reasoning.
 - The `compression:` block's `model` MUST match `auxiliary: compression: model`
-- The `compression: max_context_window: 131072` MUST match actual GPU capacity (128K)
+- The `compression: max_context_window: 131072` MUST stay at the syslog-auto pool floor (NVIDIA hosts 128K; Strix Halo 256K)
 
 ### Rule 8: GPU Workload Distribution (UPDATED 2026-07-16)
-- **RTX 3090 (24GB, 128K ctx, qwen3.6-27B-code)**: Heavy reasoning, code gen, long conversations
-- **RTX 5070 (12GB, 128K ctx, gemma-4-12b)**: Vision, web search, quick tasks, web_extract (IQ4_NL+MTP, ~65% VRAM at 128K)
-- **Strix Halo (64GB, 128K ctx, syslog-auto)**: Context compression, summarization, long docs
+- **RTX 3090 (24GB, 128K ctx, gpu-dense)**: Heavy reasoning, code gen, long conversations
+- **RTX 5070 (12GB, 128K ctx, gpu-vision)**: Vision, web search, quick tasks, web_extract (IQ4_NL+MTP, ~65% VRAM at 128K)
+- **Strix Halo (64GB, 256K ctx, syslog-auto)**: Context compression, summarization, long docs
 - Agent profiles MUST route auxiliary tasks to the correct GPU:
-  - `auxiliary.vision.model: gemma-4-12b` (RTX 5070)
-  - `auxiliary.web_extract.model: gemma-4-12b` (RTX 5070)
+  - `auxiliary.vision.model: gpu-vision` (RTX 5070)
+  - `auxiliary.web_extract.model: gpu-vision` (RTX 5070)
   - `auxiliary.compression.model: syslog-auto` (Strix Halo)
 - Default model (`model.default`) and custom_provider remain `syslog-auto` for auto-routing
 - For 128K context window: `threshold: 0.65` (fires at ~85K tokens)
 - Do NOT use `threshold: 0.25` — this fires at 65K, causing premature context loss
 - Do NOT use `threshold: 0.80` — this delays until ~105K, leaving only 23K margin
-- `max_context_window: 131072` MUST match the model's actual capacity (128K)
+- `max_context_window: 131072` MUST stay at the pool floor (NVIDIA hosts 128K; Strix Halo 256K)
 - See `devops-hermes-compression` skill for full reference
 
 ### Rule 9: Compression Threshold for 128K Models
 - For 128K context window: `threshold: 0.65` (fires at ~85K tokens)
 - Do NOT use `threshold: 0.25` — this fires at 65K, causing premature context loss
 - Do NOT use `threshold: 0.80` — this delays until ~105K, leaving only 23K margin
-- `max_context_window: 131072` MUST match the model's actual capacity (all GPUs = 128K)
+- `max_context_window: 131072` MUST stay at the pool floor (NVIDIA hosts 128K; Strix Halo 256K)
 - See `devops-hermes-compression` skill for full reference
 
 ### Rule 10: Default Model Must Be `syslog-auto` (All Agents)
 - **Koby Exception**: Per captain ruling 2026-08-11, Koby is a DeepSeek-primary external agent; its primary model remains `deepseek-v4-flash` (via api.deepseek.com to preserve DeepSeek-specific reasoning, while other sections follow Rule 10.
 - **Hermes agents**: `model.default: syslog-auto`, `custom_providers[0].model: syslog-auto`
 - **pi agents**: `defaultModel: syslog-auto` in `settings.json`, first model in `models.json`
-- `syslog-auto` is the LiteLLM routing model — it load-balances between strix-moe
-  and qwen3.6-27B-code, with gemma-4-12b as fallback. Using it protects against:
+- `syslog-auto` is the LiteLLM routing model — it load-balances across the live pool
+  (see CT 116 `/opt/inference-harness/litellm_config.yaml` for the current members and weights). Using it protects against:
   - Model name typos that cause 403 errors and silent worker failures
   - Single GPU downtime (routing falls back automatically)
   - Key/model authorization mismatches
@@ -321,7 +322,8 @@ When an agent shows "context issues" (premature compression, 401s, 504s, DeepSee
 verify ALL FOUR of these against the live config. They are the only root causes found in production:
 
 1. **max_context_window correct?** — BOTH `compression.max_context_window` AND `context.max_context_window`
-   MUST be `131072` (all GPUs are 128K). A value of `262144` causes instability near 100K and must NOT be used.
+   MUST be `131072` (the syslog-auto pool floor: NVIDIA hosts are 128K; Strix Halo is 256K).
+   A `262144` client window can route to a 128K NVIDIA host and fail, so it must NOT be used.
    ~83K instead of ~170K. Check: `grep -n max_context_window ~/.hermes/config.yaml`
 2. **base_url uses authenticated path?** — `custom_providers[0].base_url`, `delegation.base_url`,
    and ALL `auxiliary.*.base_url` MUST be `http://192.168.68.116/litellm/v1` (Rule 5, canonical)

hermes-key-enforcement.prose.md

@@ -169,16 +169,20 @@ The agent picks up the new key via `infisical run --` at gateway startup.
 
 **Keys are permanent and use bare agent name aliases.**
 
-- **Duration**: `null` — keys never expire. This is enforced by `default_key_generate_params` in `litellm_config.yaml`.
+- **Duration**: `null` — keys never expire. NOT enforced today: CT 116 `litellm_config.yaml` has no `default_key_generate_params` block, and a key generated with no explicit models comes back with an empty models list. OPEN policy question: should agent keys expire by default? (captain security-policy decision, raised separately.)
 - **Alias convention**: bare agent name only (e.g., `tanko`, `mumuni`, `koby`, `koonimo`). No dates, no versions. The alias IS the identity.
 - **Rotation triggers**: compromise, personnel departure, or quarterly security hygiene. NOT calendar-driven.
 - **Max budget**: $100 per key (config default).
 
 ```yaml
-# In litellm_config.yaml — ensures all future keys inherit these defaults:
+# NOT currently set in the authority; recommended value. CT 116 litellm_config.yaml has no
+# default_key_generate_params block today, and a key generated with no explicit models comes back
+# with an EMPTY models list. `models` is a literal key-generation parameter, so this is a value to
+# ADD — re-read the live registry at CT 116 /opt/inference-harness/litellm_config.yaml and
+# re-verify before applying.
 litellm_settings:
   default_key_generate_params:
-    models: ["syslog-auto", "qwen3.6-27B-code", "gemma-4-12b"]
+    models: ["syslog-auto", "gpu-dense", "gpu-vision", "strix-moe"]
     duration: null        # ← permanent
     max_budget: 100
     metadata:
@@ -286,13 +290,13 @@ auxiliary:
     api_key: sk-<agent-key-from-vault>           # ← workaround (get via: infisical secrets get LITELLM_API_KEY --project=agents --env=production --plain)
     api_key_env: LITELLM_API_KEY
     base_url: http://192.168.68.116/litellm/v1
-    model: gemma-4-12b
+    model: gpu-vision
     provider: harness
   compression:
     api_key: sk-<agent-key-from-vault>           # ← workaround (same as above)
     api_key_env: LITELLM_API_KEY
     base_url: http://192.168.68.116/litellm/v1
-    model: gemma-4-12b
+    model: syslog-auto
     provider: harness
 ```
 

hermes-zulip-plugin.prose.md

@@ -47,7 +47,7 @@ connectivity recovery including end-to-end DM validation.
 
 ## Requires
 
-- SSH access to target host (direct or via amdpve for CTs)
+- SSH access to target host (direct, or via the guest's Proxmox node for CTs)
 - Git repo at `https://git.sysloggh.net/SyslogSolution/zulip-platform-plugins.git`
 - Python 3 with `httpx` installed on target
 
@@ -56,7 +56,7 @@ connectivity recovery including end-to-end DM validation.
 | Host | CT | Proxmox | IP (direct) | Hermes Home | User |
 |------|-----|---------|-------------|-------------|------|
 | Tanko | CT112 | amdpve | 192.168.68.122 | /home/jerome/.hermes | jerome | *(DSH since 2026-08-27 — historical, plugin retired on this host)* |
-| Koby | CT111 | amdpve | 192.168.68.129 | /root/.hermes | root |
+| Koby | CT111 | storepve | 192.168.68.129 | /root/.hermes | root |
 | Shumba | — | — | 192.168.68.119 | /home/lucky/.hermes | lucky |
 
 | Field | Value | Trust |
@@ -72,7 +72,7 @@ connectivity recovery including end-to-end DM validation.
 ### Step 1: Resolve Target
 
 Map `target` to host, CT ID, hermes_home, and user from the live-state table.
-For CT112 and CT111, route through `ssh root@amdpve` then `pct exec <id>`.
+For CT112 route through `ssh root@amdpve`; for CT111 route through `ssh root@storepve` — then `pct exec <id>`.
 
 ### Step 2: Pull Latest Plugin Source
 

hermes-zulip-restore.prose.md

@@ -43,7 +43,7 @@ gateway restart, and connection validation.
 
 ## Requires
 
-- SSH access to target host (direct or via amdpve for CTs)
+- SSH access to target host (direct, or via the guest's Proxmox node for CTs)
 - Git repo at `https://git.sysloggh.net/SyslogSolution/zulip-platform-plugins.git`
 - Python 3 with `httpx` installed on target
 - Zulip server accessible at `https://chat.sysloggh.net`
@@ -52,7 +52,7 @@ gateway restart, and connection validation.
 
 | Host | CT | Proxmox | IP (direct) | Hermes Home | User |
 |------|-----|---------|-------------|-------------|------|
-| Koby | CT111 | amdpve | 192.168.68.129 | /root/.hermes | root |
+| Koby | CT111 | storepve | 192.168.68.129 | /root/.hermes | root |
 | Shumba | — | — | 192.168.68.119 | /home/lucky/.hermes | lucky |
 
 | Field | Value | Trust |
@@ -67,7 +67,7 @@ gateway restart, and connection validation.
 ### Step 1: Locate Target
 
 Map `target` to connectivity parameters from the live-state table above.
-For CT112 and CT111, route through `ssh root@amdpve` then `pct exec <id>`.
+For CT112 route through `ssh root@amdpve`; for CT111 route through `ssh root@storepve` — then `pct exec <id>`.
 
 ### Step 2: Deploy Zulip Adapter
 

inference-optimization.prose.md

@@ -4,7 +4,7 @@ kind: responsibility
 description: >
   Optimizes the full Syslog inference stack — LiteLLM routing weights, GPU model
   assignments, agent context management, and prompt caching — to reduce response
-  times to sub-15s average. All GPUs now at 128K context (stable ceiling).
+  times to sub-15s average. NVIDIA GPUs at 128K context; Strix Halo at 256K (2026-09-12).
 id: 067NC6KP02RG60S50M40E30928
 ---
 
@@ -23,7 +23,7 @@ management, and prompt caching — without sacrificing agent capability.
   .123, any others on .129/.122) including compression, model, context_window,
   prompt_caching, memory settings
 - `gpu-health`: health check response from all 3 GPU backends (strix-moe .15:8080,
-  qwen .8:8080, gemma .110:8080)
+  gpu-dense .8:8080, gpu-vision .110:8080)
 
 ### Maintains
 
@@ -56,14 +56,14 @@ duration.
 **Context is the root cause.** Every ~46K prompt token costs ~87s of
 prefill time at 532 tok/s. Fix context first, routing second.
 
-- **Route by task**: qwen for code/standard queries; gemma for
-  compression/auxiliary; strix-moe for compression tasks.
+- **Route by task**: gpu-dense for code/standard queries; gpu-vision for
+  vision/web-auxiliary; syslog-auto for compression.
 - **Compress aggressively**: threshold at 40% (not 65%) — a 128K window should
   compact at 51K, not 85K. Target 15% tail (not 30%).
 - **Cache everything repeated**: system prompts, skill docs, AGENTS.md — these
   never change between turns. Single-digit cache hit rate is unacceptable.
 - **Lower context ceiling**: 128K window is the stable ceiling for agent conversations.
-  GPUs reduced from 256K to 128K (2026-07-17). 128K window should compact at 85K (0.65 threshold). For larger contexts, route to external providers.
+  GPUs reduced from 256K to 128K (2026-07-17) for the NVIDIA hosts; Strix Halo runs 256K (2026-09-12). 128K window should compact at 85K (0.65 threshold). For larger contexts, route to external providers.
 
 ### Shape
 
@@ -96,8 +96,10 @@ call enable-prompt-caching
   hosts: [192.168.68.15, 192.168.68.8, 192.168.68.110]
 
 -- Phase 5: Verify end-to-end latency
+-- `models` is a literal verification parameter (a snapshot only): the authoritative registry is
+-- CT 116 /opt/inference-harness/litellm_config.yaml; re-read it before use.
 
 call verify-latency
   host: 192.168.68.116
-  models: [syslog-auto, qwen3.6-27B-code, gemma-4-12b]
+  models: [syslog-auto, gpu-dense, gpu-vision, strix-moe]
 ```

infrastructure-control.prose.md

@@ -16,8 +16,8 @@ description: >
   **Last verified:** 2026-08-15 — hwepve removed from Tabiri cluster
   (now 5 nodes: minipve, amdpve, storepve, acerpve, ocupve). hwepve
   (192.168.68.4) is a standalone PVE node + NetBird routing peer;
-  London relocation pending. CTs 100 (abiba) and 105 (kagentz) moved
-  to minipve.
+  London relocation pending. CT 100 (abiba) is on minipve; CT 105
+  (kagentz) is on amdpve.
 ---
 
 # Infrastructure Control Pattern
@@ -105,16 +105,16 @@ description: >
 
 | Node | IP | CPU | RAM | VMs/CTs | Role |
 |------|----|-----|-----|---------|------|
-| minipve | .12 | 16C | 30GB | abiba, kagentz, authentik, gitea, syslog-api, infisical-vault, jitsi | Auth, git, messaging |
-| amdpve | .15 | 32C | 62GB | tanko, tdunna, baggy, scottdenya | Agents, compute |
-| storepve | .6 | 28C | 31GB | docker-vm, ra-h-os, PBS, media, jdownloader, zulip | Docker, storage, chat |
+| minipve | .12 | 16C | 30GB | abiba, authentik, gitea, syslog-api, infisical-vault, jitsi | Auth, git, messaging |
+| amdpve | .15 | 32C | 62GB | kagentz, tanko, baggy, scottdenya, adguard2 | Agents, compute |
+| storepve | .6 | 28C | 31GB | docker-vm, ra-h-os, PBS, media, jdownloader, zulip, tdunna | Docker, storage, chat |
 | acerpve | .9 | 28C | 31GB | llm-gpu | GPU VMs |
 | ocupve | .5 | 12C | 14GB | ocu-llm | GPU VMs |
 
-> **Note:** CTs on storepve include jdownloader (CT 118). AdGuard (CT 102) is on
-> minipve at .10, not acerpve. Abiba (CT 100) and kagentz (CT 105) are on
-> minipve (moved from hwepve 2026-08-15). Mumuni runs inside Abiba CT100
-> (.24); CT 114 (mumuni) no longer exists in the cluster.
+> **Note:** CTs on storepve include jdownloader (CT 118) and tdunna (CT 111).
+> AdGuard (CT 102) is on minipve at .10, not acerpve. Abiba (CT 100) is on
+> minipve (moved from hwepve 2026-08-15); kagentz (CT 105) is on amdpve.
+> Mumuni runs inside Abiba CT100 (.24); CT 114 (mumuni) no longer exists in the cluster.
 >
 > **hwepve (192.168.68.4) — STANDALONE (removed from Tabiri 2026-08-15):**
 > Huawei MateBook 16 (KLVL-WXX9), pve-manager/9.2.10, kernel 7.0.14-8-pve.
@@ -222,8 +222,8 @@ description: >
 
 **Prometheus targets**:
 - 192.168.68.8:9400 (RTX 3090 — qwen)
-- 192.168.68.110:9400 (RTX 5070 — gemma)
-- 192.168.68.15:9400 (Strix Halo — qwen3.6-35B-udq4)
+- 192.168.68.110:9400 (RTX 5070 — gpu-vision)
+- 192.168.68.15:9400 (Strix Halo — strix-moe)
 - harness-litellm:4000 (LiteLLM health)
 
 ### Ecosystem C: Netbird (72.61.0.17 — Hostinger srv1079750.hstgr.cloud)
@@ -602,13 +602,13 @@ ssh root@192.168.68.110 "systemctl restart llama-server"
 | 102 | adguard | **minipve** | **.10** | DNS | ❌ |
 | 103 | ocu-llm | ocupve | .110 | GPU RTX 5070 | ❌ |
 | 104 | authentik | minipve | .11 | OIDC | ❌ |
-| 105 | kagentz | minipve | — | Agent Zero | ✅ |
+| 105 | kagentz | amdpve | — | Agent Zero | ✅ |
 | 106 | ra-h-os | storepve | .65 | KG bridge | ✅ MCP |
 | 107 | pbs | storepve | — | Backups | ❌ |
 | 108 | media | storepve | — | Media | ❌ |
 | 109 | docker-vm | storepve | .7 | Docker host | ❌ |
 | 110 | gitea | minipve | **.17** | Git | ❌ |
-| 111 | tdunna | amdpve | .129 | Hermes agent | ✅ |
+| 111 | tdunna | storepve | .129 | Hermes agent — ⛔ REPORT-ONLY (Theo's box, no GC) | ✅ |
 | 112 | tanko | amdpve | .122 | DSH (DeepSeek Harness) agent | ✅ |
 | 113 | baggy | amdpve | .114 | Hermes agent | ✅ |
 | 115 | scottdenya | amdpve | .75 | Denya OneCare | ❌ |
@@ -616,6 +616,7 @@ ssh root@192.168.68.110 "systemctl restart llama-server"
 | 117 | zulip | storepve | .19 | Chat | ❌ |
 | 118 | jdownloader | storepve | .20 | JDownloader LXC (dedicated, migrated from docker-vm 2026-08-01) | ✅ |
 | 119 | infisical-vault | minipve | — | Vault | ❌ |
+| 120 | adguard2 | amdpve | — | DNS (secondary AdGuard) | ❌ |
 
 ## Appendix C: Docker Compose Files Location
 
@@ -636,8 +637,8 @@ Source of truth: `/root/scripts/pct-run.sh` or `prose-contracts/scripts/pct-run.
 | CT | Name | Node | pct-run |
 |-----|------|------|---------|
 | 100 | abiba | minipve | `pct-run 100` |
-| 105 | kagentz | minipve | `pct-run 105` |
-| 111 | tdunna | amdpve | `pct-run 111` |
+| 105 | kagentz | amdpve | `pct-run 105` |
+| 111 | tdunna | storepve | `pct-run 111` (⛔ report-only — no GC) |
 | 112 | tanko | amdpve | `pct-run 112` |
 | 113 | baggy | amdpve | `pct-run 113` |
 | 115 | scottdenya | amdpve | `pct-run 115` |
@@ -648,6 +649,9 @@ Source of truth: `/root/scripts/pct-run.sh` or `prose-contracts/scripts/pct-run.
 | 107 | proxmox-backup | storepve | `pct-run 107` |
 | 108 | media | storepve | `pct-run 108` |
 | 117 | zulip | storepve | `pct-run 117` |
+| 118 | jdownloader | storepve | `pct-run 118` |
+| 119 | infisical-vault | minipve | `pct-run 119` |
+| 120 | adguard2 | amdpve | `pct-run 120` |
 | 102 | adguard | **minipve** | `pct-run 102` |
 
 GPU bare-metal hosts (.8 acerpve, .110 ocupve, .15 amdpve) are NOT CTs — use SSH directly:

infrastructure-monitoring.prose.md

@@ -129,10 +129,15 @@ the any-HTTP rule. On those — the authenticated Zulip POST and the router
 pwd -P
 
 # Zulip API health (POST ping)
-source /etc/litellm-monitor.env
-ZULIP_USER="abiba-bot@chat.sysloggh.net"
-curl -s -o /dev/null -w '%{http_code}' -X POST https://chat.sysloggh.net/api/v1/messages -u "${ZULIP_USER}:${ZULIP_BOT_KEY}"
-# Expected: 200 (HTTP 000 = unreachable/cache)
+# NOTE: /etc/litellm-monitor.env exists only on CT 116, retrieve keys from CT 116 via:
+zulip_key=$(ssh root@192.168.68.116 "grep ZULIP_BOT_KEY /etc/litellm-monitor.env | cut -d= -f2")
+if [ -z "$zulip_key" ]; then
+  echo "credential-missing: ZULIP_BOT_KEY not found in /etc/litellm-monitor.env"
+else
+  ZULIP_USER="abiba-bot@chat.sysloggh.net"
+  curl -s -o /dev/null -w '%{http_code}' -X POST https://chat.sysloggh.net/api/v1/messages -u "${ZULIP_USER}:${zulip_key}"
+  # Expected: 200 (HTTP 000 = unreachable/cache)
+fi
 
 # PM2 process health
 pm2 jlist

litellm-api-keys.prose.md

@@ -67,8 +67,10 @@ description: >
 4. **If action == "create"**:
    - Generate new key with key_alias: "{agent_name}" (e.g., "tanko" — bare name, no date)
    - Set metadata: { "agent": "{agent_name}", "purpose": "agent-inference" }
-   - Duration is null (permanent) — inherited from litellm default_key_generate_params
-   - Set models: ["syslog-auto", "qwen3.6-27B-code", "gemma-4-12b", "strix-moe", "gpu-dense", "gpu-light", "qwen3.6-35B-udq4"]
+   - Duration is whatever the caller passes; NO default enforcement exists today (CT 116 `litellm_config.yaml` has no `default_key_generate_params` block, and a key with no explicit models returns an empty models list). Agent keys are permanent by policy, not by that block. OPEN policy question: should agent keys expire by default? (captain security-policy decision, raised separately.)
+   - Set models: read the live key-scoped set rather than hardcoding one — `/v1/models` is key-scoped,
+     and the authoritative registry is CT 116 `/opt/inference-harness/litellm_config.yaml`. Do not add
+     retired names (`gemma-4-12b`, `gpu-light`, `crew-auto` — all retired 2026-09-12).
    - Note: `ornith-1.0-35b` is NOT a valid LiteLLM model name (use `strix-moe`, the stable alias). qwen3.6-35B-A3B removed from fleet (was never deployed).
    - Return the new key
 5. **If action == "rotate"**:

litellm-client-timeouts.prose.md

@@ -26,9 +26,8 @@ description: >
 | Model | avg latency | avg TTFT | p-profile (24h) |
 |---|---|---|---|
 | syslog-auto | 28.8s | 25.4s | 68 calls took 30-120s; tail to ~300s under load |
-| qwen3.6-27B-code | 23.0s | — | same backend class as syslog-auto |
 | strix-moe | 7.5s | — | Strix Halo, healthy |
-| gemma-4-12b | 2.6s | — | RTX 5070, healthy |
+| gpu-vision (retired gemma-4-12b, RTX 5070) | 2.6s | — | RTX 5070, healthy |
 
 Sep 6 incident timeline: failures 04:00-07:00 EDT (0% GPU util = wedged
 backend), full recovery 07:00-08:00 with ZERO client failures once requests
@@ -53,13 +52,11 @@ proxy queuing.
 
 ### 2. Auxiliary tasks — keep template timeouts, one correction
 
-- vision: 60s (keep), web_extract: 30s (keep) — gemma-4-12b averages 2.6s;
-  these are fine.
+- vision: 60s (keep), web_extract: 30s (keep) — the 2.6s average was measured on `gemma-4-12b` (retired 2026-09-12); the live RTX 5070 alias is `gpu-vision`.
 - compression: 300s (keep — this was already raised from 60 per gpu-fleet).
 - **gpu-dense delegation/x_search: set timeout >= 120s.** The RTX 3090
-  (qwen3.6-27B-code backend, 23.0s avg) is the same speed class as
-  syslog-auto; delegation defaults that assume fast responses will 408 the
-  same way.
+  (gpu-dense backend) is the same speed class as syslog-auto; delegation
+  defaults that assume fast responses will 408 the same way.
 
 ### 3. Retry policy — backoff, not repetition
 

litellm-health.prose.md

@@ -44,7 +44,7 @@ Request → nginx:80 → LiteLLM:4000 → GPU(llama-server, parallel 2)
 **What changed (v3.2.0 → v4.0.0 — 2026-07-08)**:
 - Router REMOVED from request path — LiteLLM proxies directly to GPU
 - All GPUs at parallel 2 (was parallel 1)
-- NVIDIA context reduced 256K→128K to free VRAM — now the stable ceiling across all GPUs (2026-07-17)
+- NVIDIA context reduced 256K→128K to free VRAM — the stable NVIDIA ceiling (2026-07-17); Strix Halo runs 256K (2026-09-12)
 - Timeouts and fallback chains are config state — read them from CT 116 `/opt/inference-harness/litellm_config.yaml`; they are not duplicated here.
 
 ## Parameters
@@ -69,7 +69,16 @@ Request → nginx:80 → LiteLLM:4000 → GPU(llama-server, parallel 2)
 - Network access to public_url, auth_host, and gpu_dashboard_url
 - LiteLLM master key for admin endpoints only (`/key/list`, `/key/generate`, `/key/info`)
 - The dedicated `monitor` agent key on CT 116 at `/etc/litellm-monitor.env` (root-only 0600) for
-  model inference checks — the master key must never be used for inference
+  model inference checks, scoped for every alias step 7 probes (`gpu-dense`, `gpu-vision`,
+  `strix-moe`, `syslog-auto`). Retrieve from the executor's host via:
+
+  ```
+  monitor key: ssh root@192.168.68.116 "grep LITELLM_MONITOR_KEY /etc/litellm-monitor.env | cut -d= -f2"
+  master key:  ssh root@192.168.68.116 "docker exec harness-litellm printenv LITELLM_MASTER_KEY"
+  ```
+
+  If credentials are missing or unreadable, the probe must report `credential-missing` (not bare 401 or "0 keys").
+  The master key must never be used for inference.
 
 ## GPU Fleet Topology
 
@@ -144,12 +153,19 @@ contracts — read them there. Do not re-add retired names (`gemma-4-12b`, `gpu-
 7. **Check model inference via LiteLLM** — Test one model on each GPU host. The health
    check runs on the **backend edge**, not the public edge, so these paths carry the
    `/litellm/` prefix:
-   - POST http://{{backend_host}}/litellm/v1/chat/completions model=qwen3.6-27B-code → expect 200 (RTX 3090, .8)
+   - POST http://{{backend_host}}/litellm/v1/chat/completions model=gpu-dense → expect 200 (RTX 3090, .8)
    - POST http://{{backend_host}}/litellm/v1/chat/completions model=gpu-vision → expect 200 (RTX 5070, .110)
    - POST http://{{backend_host}}/litellm/v1/chat/completions model=strix-moe → expect 200 (Strix Halo, .15)
-   - Auth uses the dedicated `monitor` agent key, read on CT 116 from
-     `/etc/litellm-monitor.env` (root-only 0600). Do NOT use the master key for inference —
-     the master key is for admin endpoints only (`/key/list`, `/key/generate`, `/key/info`).
+   - Auth uses the dedicated `monitor` agent key. Retrieve via:
+     `ssh root@192.168.68.116 "grep LITELLM_MONITOR_KEY /etc/litellm-monitor.env | cut -d= -f2"`
+     Do NOT use the master key for inference — the master key is for admin endpoints only
+     (`/key/list`, `/key/generate`, `/key/info`). Retrieve master key via:
+     `ssh root@192.168.68.116 "docker exec harness-litellm printenv LITELLM_MASTER_KEY"`
+   - KEY SCOPE: the `monitor` key MUST be scoped for the three probed aliases (`gpu-dense`,
+     `gpu-vision`, `strix-moe`) plus the `syslog-auto` fallback pool, otherwise the probe
+     returns 403 and the host is not covered.
+     If a probe returns 403, widen the monitor key's model list on CT 116 (add the missing
+     alias) and re-run — never drop the host from the probe to make the check pass.
    - `gemma-4-12b` was retired and returns 400 `Invalid model name` — do not re-add it to
      this list. The RTX 5070 host now serves `gpu-vision`.
    - `/v1/models` is **key-scoped**: a model is only visible to keys allowed to use it, so
@@ -161,6 +177,9 @@ contracts — read them there. Do not re-add retired names (`gemma-4-12b`, `gpu-
 
 8. **Check agent keys**:
    - GET http://{{backend_host}}/litellm/key/list with master key (admin endpoint) → verify all 6 agents have keys
+   - **IMPORTANT**: Run the curl on the CT 116 HOST, not inside the container. The `harness-litellm` container has no curl/wget. Use:
+     `ssh root@192.168.68.116 "curl -s -H 'Authorization: Bearer $MASTER_KEY' http://127.0.0.1:4000/key/list"`
+   - If the response is empty or unparseable, report `admin-call-failed` (not "0 agent keys")
 
 9. **Check Grafana**:
    - GET {{grafana_url}}/api/health → expect 200

mumuni-delegation-prose-contract.prose.md

@@ -90,8 +90,8 @@ raw data never provided.
 
 | Worker | Model | Toolsets | Role | Use When |
 |--------|-------|----------|------|----------|
-| `syslog-code` | qwen3.6-27B-code | terminal, file, web, memory, skills | Code patches, automation, scripts | Writing/modifying code, creating scripts, debugging, reading/writing files |
-| `syslog-devops` | qwen3.6-27B-code | terminal, file, web, memory, skills | Infrastructure, DB, bridge, Proxmox | Server ops, SSH, Docker, Proxmox, DB queries, hardware checks |
+| `syslog-code` | gpu-dense | terminal, file, web, memory, skills | Code patches, automation, scripts | Writing/modifying code, creating scripts, debugging, reading/writing files |
+| `syslog-devops` | gpu-dense | terminal, file, web, memory, skills | Infrastructure, DB, bridge, Proxmox | Server ops, SSH, Docker, Proxmox, DB queries, hardware checks |
 | `syslog-email` | strix-moe | terminal, file, web, memory, skills | Email automation, mail operations | Sending/receiving email, inbox management, SMTP operations |
 | `syslog-research` | strix-moe | terminal, file, web, memory, skills, **browser** | Analysis, classification, data processing | Web research, browser tasks, data analysis, classification, reading docs |
 | `syslog-review` | strix-moe | terminal, file, web, memory, skills | Verification, QA, audit validation | **ALWAYS** verify worker output before delivery — especially for infra changes, code builds, and research findings |

pm2-self-heal.prose.md

@@ -9,7 +9,6 @@ description: >
 - abiba-telegram: { status: "online", uptime: string, restarts: number }
 - abiba-zulip: { status: "online", uptime: string, restarts: number }
 - gitea-runner: { status: "online", uptime: string, restarts: number }
-- spoton-service: { status: "online", uptime: string, restarts: number }
 - zulip-watchdog: { status: "online", uptime: string, restarts: number }
 - last_check: timestamp
 

proxmox-monitor.prose.md

@@ -87,7 +87,7 @@ agent: abiba
 | storepve | 192.168.68.6 | PVE |
 | acerpve | 192.168.68.9 | PVE (hosts llm-gpu qemu/101) |
 | minipve | 192.168.68.12 | PVE |
-| amdpve | 192.168.68.15 | PVE + Strix Halo LLM (qwen3.6-35B-udq4, strix-moe) |
+| amdpve | 192.168.68.15 | PVE + Strix Halo LLM (strix-moe) |
 
 ## Operations
 

runs/disk-gc-ct111-report-only-verification-20260912.log

@@ -0,0 +1,44 @@
+disk-gc report-only verification — CT 111 / tdunna / 192.168.68.129
+date: 2026-09-12T19:11:36Z
+host: abiba (this scanner runs INSIDE CT 100 / abiba)
+branch head: 6e612ce37b1b9a9688b04e7f816848e7a4185fca
+command: scripts/disk-gc-plan.py --scan <live fleet scan>
+
+PURPOSE: prove that on a REAL fleet scan, CT 111 is alerted and NO gc-executor action
+is emitted for it at any level. No GC command was executed against .129.
+
+--- live fleet scan (df -P / via scripts/pct-run.sh for LXC, direct SSH for hosts) ---
+  tdunna                                    84%
+  acerpve                  192.168.68.9     77%
+  amdpve                   192.168.68.15    76%
+  ocu-llm                  192.168.68.110   69%
+  storepve                 192.168.68.6     65%
+  kagentz                                   61%
+  tanko                                     56%
+  minipve                  192.168.68.12    49%
+  authentik                                 45%
+  infisical-vault                           39%
+  adguard                                   38%
+  ocupve                   192.168.68.5     38%
+  scottdenya                                35%
+  syslog-api                                34%
+  llm-gpu                  192.168.68.8     25%
+  abiba                                     23%
+  baggy                                     22%
+  jdownloader                               21%
+  gitea                                     16%
+  ra-h-os                                   13%
+  zulip                                     12%
+  docker-vm                192.168.68.7     11%
+  adguard2                                  10%
+  media                                     9%
+  proxmox-backup-server                     4%
+
+--- planner output (action plan) ---
+    111 AMBER 84.0% -> REPORT-ONLY (no GC) — Theo's box — captain ruling 2026-08-17, re-confirmed 2026-09-10
+    acerpve AMBER 77.0% -> gc-executor
+    amdpve AMBER 76.0% -> gc-executor
+
+--- verdict ---
+  CT 111 (tdunna) 84% AMBER -> report-only; no gc-executor row emitted; no GC run on .129.
+  Owned hosts acerpve .9 (77%) and amdpve .15 (76%) -> gc-executor (ours).

scripts/disk-gc-plan.py

@@ -0,0 +1,206 @@
+#!/usr/bin/env python3
+"""disk-gc-plan — turn a fleet disk scan into the GC action plan.
+
+This is the executable side of `disk-gc-threat-response.prose.md`. It exists so the
+report-only gate is enforced by code that can be tested, rather than by prose the
+executor might misread.
+
+THE HARD GATE: guests listed in the contract's `report_only_guests` block are
+DETECT-AND-REPORT-ONLY at EVERY level (AMBER, RED, CRITICAL). This tool will never
+emit a `gc-executor` action for one, so no GC command can be constructed for it.
+
+The gate is keyed on GUEST identity — guest id, hostname, or IP — never on an agent
+name. An agent-name marker can silently miss the guest it lives on; a guest marker
+cannot.
+
+The authoritative exclusion list lives in the contract itself (the fenced ```yaml
+block containing `report_only_guests:`). This tool reads it from there so there is
+only ever one copy.
+
+Usage:
+    disk-gc-plan.py --scan scan.json            # [{"id":111,"usage_pct":84}, ...]
+    cat scan.json | disk-gc-plan.py             # same, via stdin
+    disk-gc-plan.py --scan scan.json --json     # machine-readable plan
+
+Scan entries may carry any of: id / guest / vmid / ct / ctid, hostname / name, ip.
+A threshold-crossing entry with no recognizable identity is reported, never GC'd.
+Exit codes: 0 ok, 1 usage/parse error.
+"""
+from __future__ import annotations
+
+import argparse
+import json
+import pathlib
+import re
+import sys
+
+import yaml
+
+REPO = pathlib.Path(__file__).resolve().parent.parent
+DEFAULT_CONTRACT = REPO / "disk-gc-threat-response.prose.md"
+
+AMBER, RED, CRITICAL = 75, 85, 95
+
+IDENTITY_FIELDS = ("id", "guest", "vmid", "ct", "ctid", "hostname", "name", "ip")
+IDENTITY_TYPE_PREFIX = re.compile(r"^(?:lxc|qemu)/")
+UNIDENTIFIED_REASON = "unidentified target - refusing to schedule GC"
+
+
+def load_report_only_guests(contract_path: pathlib.Path) -> list[dict]:
+    """Read the authoritative report_only_guests block out of the contract.
+
+    The contract carries it as a fenced ```yaml block. Parsing the declared,
+    machine-readable block is the intended interface — the contract owns the list.
+    """
+    text = contract_path.read_text(encoding="utf-8")
+    for block in re.findall(r"```yaml\n(.*?)```", text, re.S):
+        if "report_only_guests:" in block:
+            data = yaml.safe_load(block)
+            guests = data.get("report_only_guests") or []
+            if not isinstance(guests, list):
+                raise SystemExit("report_only_guests must be a list")
+            if not guests:
+                raise SystemExit(
+                    "report_only_guests is empty or missing - refusing to plan GC "
+                    "without the report-only gate"
+                )
+            for guest in guests:
+                if not isinstance(guest, dict) or not _keys(guest):
+                    raise SystemExit(
+                        "report_only_guests entry has no recognizable identity key "
+                        f"(expected one of: {', '.join(IDENTITY_FIELDS)}): {guest!r}"
+                    )
+            return guests
+    raise SystemExit(
+        f"no authoritative report_only_guests block found in {contract_path}"
+    )
+
+
+def _canonical_number(number: float) -> str:
+    if float(number).is_integer():
+        return str(int(number))
+    return str(number).strip().lower()
+
+
+def _normalize_identity(value: object) -> str:
+    """Canonicalise a guest identity so differently-encoded ids compare equal:
+    numeric and numeric-string ids collapse to an integer string, Proxmox
+    type prefixes and leading zeros are stripped, and hostnames/IPs are only
+    trimmed and lowercased."""
+    if isinstance(value, bool):
+        return str(value).strip().lower()
+    if isinstance(value, (int, float)):
+        return _canonical_number(float(value))
+    text = str(value).strip().lower()
+    text = IDENTITY_TYPE_PREFIX.sub("", text)
+    try:
+        return _canonical_number(float(text))
+    except ValueError:
+        return text
+
+
+def _keys(entry: dict) -> set[str]:
+    """Guest/host identity keys, shared by exclusions and scan entries so the two
+    sides of the gate can never key on different fields."""
+    out: set[str] = set()
+    for field in IDENTITY_FIELDS:
+        value = entry.get(field)
+        if value is None:
+            continue
+        key = _normalize_identity(value)
+        if key:
+            out.add(key)
+    return out
+
+
+def level_for(pct: float) -> str | None:
+    if pct >= CRITICAL:
+        return "CRITICAL"
+    if pct >= RED:
+        return "RED"
+    if pct >= AMBER:
+        return "AMBER"
+    return None
+
+
+def build_plan(scan: list[dict], report_only: list[dict]) -> list[dict]:
+    excluded = [(e, _keys(e)) for e in report_only]
+    plan: list[dict] = []
+    for entry in scan:
+        pct = entry.get("usage_pct")
+        if pct is None:
+            continue
+        level = level_for(float(pct))
+        if level is None:
+            continue  # GREEN: log only, no action
+        scan_keys = _keys(entry)
+        target = next(
+            (entry.get(k) for k in IDENTITY_FIELDS if entry.get(k) not in (None, "")),
+            "?",
+        )
+        if not scan_keys:
+            plan.append({
+                "target": target,
+                "level": level,
+                "pct": float(pct),
+                "action": "report-only",
+                "reason": UNIDENTIFIED_REASON,
+            })
+            continue
+        match = next((e for e, keys in excluded if keys & scan_keys), None)
+        if match is not None:
+            plan.append({
+                "target": target,
+                "level": level,
+                "pct": float(pct),
+                "action": "report-only",
+                "reason": match.get("reason", "").strip(),
+            })
+        else:
+            plan.append({
+                "target": target,
+                "level": level,
+                "pct": float(pct),
+                "action": "gc-executor",
+            })
+    plan.sort(key=lambda row: row["pct"], reverse=True)
+    return plan
+
+
+def main() -> int:
+    ap = argparse.ArgumentParser(description="Plan disk GC actions with the report-only gate.")
+    ap.add_argument("--scan", help="JSON file: list of {id|ct|hostname|ip, usage_pct}")
+    ap.add_argument("--contract", default=str(DEFAULT_CONTRACT))
+    ap.add_argument("--json", action="store_true", help="emit the plan as JSON")
+    args = ap.parse_args()
+
+    raw = pathlib.Path(args.scan).read_text() if args.scan else sys.stdin.read()
+    try:
+        scan = json.loads(raw)
+    except json.JSONDecodeError as exc:
+        print(f"invalid scan JSON: {exc}", file=sys.stderr)
+        return 1
+    if not isinstance(scan, list):
+        print("scan must be a JSON list", file=sys.stderr)
+        return 1
+
+    report_only = load_report_only_guests(pathlib.Path(args.contract))
+    plan = build_plan(scan, report_only)
+
+    if args.json:
+        print(json.dumps(plan, indent=2))
+        return 0
+
+    if not plan:
+        print("no threats (nothing at or above 75%)")
+        return 0
+    for row in plan:
+        if row["action"] == "report-only":
+            print(f"  {row['target']} {row['level']} {row['pct']}% -> REPORT-ONLY (no GC) — {row['reason']}")
+        else:
+            print(f"  {row['target']} {row['level']} {row['pct']}% -> gc-executor")
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())

scripts/pct-run.sh

@@ -11,11 +11,13 @@ set -euo pipefail
 # ── CT ID → PVE Node mapping (maintained HERE, not in prose contracts) ──
 declare -A CT_NODES=(
     # amdpve (192.168.68.15)
-    [111]=amdpve    # tdunna
+    [105]=amdpve    # kagentz (was hwepve — corrected 2026-09-12; live per pvesh)
     [112]=amdpve    # tanko
     [113]=amdpve    # baggy
     [115]=amdpve    # scottdenya
+    [120]=amdpve    # adguard2 (added 2026-09-12)
     # minipve (192.168.68.12)
+    [100]=minipve   # abiba (was hwepve)
     [102]=minipve   # adguard (was acerpve)
     [104]=minipve   # authentik
     [110]=minipve   # gitea
@@ -25,17 +27,16 @@ declare -A CT_NODES=(
     [106]=storepve  # ra-h-os
     [107]=storepve  # proxmox-backup
     [108]=storepve  # media
+    [111]=storepve  # tdunna (was amdpve — corrected 2026-09-12; live per pvesh)
     [117]=storepve  # zulip
     [118]=storepve  # jdownloader
     # acerpve (192.168.68.9) — no CTs (bare metal GPU .8)
-    [100]=minipve   # abiba (was hwepve)
-    [105]=minipve   # kagentz (was hwepve)
     # ocupve (192.168.68.5) — no CTs (bare metal GPU .110)
     #
     # REMOVED CTs (migrated to bare metal, decommissioned, or VMs):
-    #   101 llm-gpu   → bare metal 192.168.68.8   (RTX 3090)
-    #   103 ocu-llm   → bare metal 192.168.68.110 (RTX 5070)
-    #   109 docker-vm → KVM VM 192.168.68.7       (use direct SSH)
+    #   101 llm-gpu   → bare metal 192.168.68.8   (RTX 3090)  [QEMU VM on acerpve]
+    #   103 ocu-llm   → bare metal 192.168.68.110 (RTX 5070)  [QEMU VM on ocupve]
+    #   109 docker-vm → KVM VM 192.168.68.7       (use direct SSH) [QEMU VM on storepve]
 )
 
 # Each node must be root-accessible via SSH hostname

scripts/prose-ai-review.sh

@@ -47,17 +47,19 @@ You are a code reviewer for OpenProse infrastructure contracts in the Syslog Sol
 The infrastructure-control.prose.md contract is the canonical reference for the cluster topology:
 
 **Proxmox Cluster "Tabiri" (5 nodes):**
-- amdpve (192.168.68.15): tanko, tdunna, baggy, scottdenya
-- minipve (192.168.68.12): abiba, kagentz, adguard, authentik, gitea, syslog-api, infisical-vault
-- storepve (192.168.68.6): docker-vm, ra-h-os, PBS, media, jdownloader, zulip
+- amdpve (192.168.68.15): kagentz, tanko, baggy, scottdenya, adguard2
+- minipve (192.168.68.12): abiba, adguard, authentik, gitea, syslog-api, infisical-vault
+- storepve (192.168.68.6): docker-vm, ra-h-os, PBS, media, jdownloader, zulip, tdunna
 - acerpve (192.168.68.9): llm-gpu
 - ocupve (192.168.68.5): ocu-llm
 
-**CT IDs (verified 2026-07-24 against PVE API):**
+**CT IDs (verified 2026-09-12 against PVE API):**
 100:abiba  102:adguard  104:authentik  105:kagentz  106:ra-h-os
 107:pbs    108:media    110:gitea      111:tdunna   112:tanko
 113:baggy  115:scottdenya  116:syslog-api 117:zulip
-118:jdownloader  119:infisical-vault
+118:jdownloader  119:infisical-vault  120:adguard2
+
+**CT 111 (tdunna, 192.168.68.129) is REPORT-ONLY — Theo's box; alert only, never garbage-collect.**
 
 **NO CT 122, CT 123, or .19 exist in the cluster.**
 

tests/test_audit_hermes_config_alias.py

@@ -0,0 +1,191 @@
+"""Regression tests for the 2026-09-12 retired-alias sweep in audit-hermes-config.py.
+
+WHY THIS FILE EXISTS: the executable audit pushed agent configs toward a DEAD alias.
+Rule 8 required `auxiliary.vision.model == "gpu-light"` and
+`auxiliary.web_extract.model == "gpu-light"`, but `gpu-light` (and its raw predecessor
+`gemma-4-12b`) were retired on 2026-09-12 and now return 400 `Invalid model name`; the
+live RTX 5070 alias is `gpu-vision`. A config that adopted the correct canonical alias
+therefore FAILED our own audit, so the audit was actively enforcing a broken config.
+
+These tests execute the real CLI (`python3 audit-hermes-config.py <config>`) and assert
+observable behaviour — exit code and the emitted rule message — for the live alias and
+for both retired names. No network, vault, or SSH access is required.
+"""
+from __future__ import annotations
+
+import pathlib
+import subprocess
+import sys
+
+ROOT = pathlib.Path(__file__).resolve().parent.parent
+AUDIT = ROOT / "audit-hermes-config.py"
+
+BASE = """
+model:
+  api_key: ""
+  api_key_env: LITELLM_API_KEY
+  base_url: http://192.168.68.116/v1
+  max_tokens: 4096
+  default: syslog-auto
+  provider: harness
+fallback_providers:
+  provider: deepseek
+  model: deepseek-v4-flash
+  api_key_env: DEEPSEEK_API_KEY
+compression:
+  model: syslog-auto
+  provider: harness
+  threshold: 0.65
+  max_context_window: 131072
+auxiliary:
+  vision:
+    model: {alias}
+    provider: harness
+  web_extract:
+    model: {alias}
+    provider: harness
+  compression:
+    model: syslog-auto
+    provider: harness
+delegation:
+  provider: harness
+custom_providers:
+  - name: harness
+    key_env: LITELLM_API_KEY
+    base_url: http://192.168.68.116/v1
+"""
+
+
+def _run_config(tmp_path, name, text):
+    cfg = tmp_path / name
+    cfg.write_text(text)
+    proc = subprocess.run(
+        [sys.executable, str(AUDIT), str(cfg)],
+        capture_output=True, text=True,
+    )
+    return proc.returncode, proc.stdout
+
+
+def _run(tmp_path, alias):
+    return _run_config(tmp_path, f"{alias}.yaml", BASE.format(alias=alias))
+
+
+def test_live_canonical_alias_passes(tmp_path):
+    """The RTX 5070 alias that actually resolves must satisfy Rule 8."""
+    code, out = _run(tmp_path, "gpu-vision")
+    assert code == 0, out
+    assert "RESULT: PASS" in out
+
+
+def test_retired_gpu_light_is_rejected(tmp_path):
+    """A config pinned to the retired alias must fail, not pass."""
+    code, out = _run(tmp_path, "gpu-light")
+    assert code == 1, out
+    assert "auxiliary.vision.model must be gpu-vision" in out
+    assert "RESULT: FAIL" in out
+
+
+def test_retired_gemma_is_rejected(tmp_path):
+    """The retired raw model name must fail Rule 8 as well."""
+    code, out = _run(tmp_path, "gemma-4-12b")
+    assert code == 1, out
+    assert "auxiliary.vision.model must be gpu-vision" in out
+    assert "RESULT: FAIL" in out
+
+
+def test_corrected_compression_example_passes(tmp_path):
+    """The corrected workaround (vision=gpu-vision, compression=syslog-auto) must PASS."""
+    code, out = _run(tmp_path, "gpu-vision")
+    assert code == 0, out
+    assert "[Rule 7] compression.model must be syslog-auto (got 'syslog-auto')" in out
+    assert "[Rule 7] auxiliary.compression.model must be syslog-auto (got 'syslog-auto')" in out
+    assert "RESULT: PASS" in out
+
+
+def test_retired_alias_in_delegation_is_rejected(tmp_path):
+    """delegation.model has no dedicated value rule, so a retired name there used to PASS."""
+    code, out = _run_config(
+        tmp_path,
+        "delegation-gpu-light.yaml",
+        BASE.format(alias="gpu-vision").replace(
+            "delegation:\n  provider: harness",
+            "delegation:\n  provider: harness\n  model: gpu-light",
+        ),
+    )
+    assert code == 1, out
+    assert "delegation.model = 'gpu-light' is retired" in out
+    assert "RESULT: FAIL" in out
+
+
+def test_retired_alias_in_custom_providers_is_rejected(tmp_path):
+    """custom_providers[*].model is model-bearing; a retired name there must fail."""
+    code, out = _run_config(
+        tmp_path,
+        "custom-provider-gpu-light.yaml",
+        BASE.format(alias="gpu-vision").replace(
+            "  - name: harness\n    key_env: LITELLM_API_KEY",
+            "  - name: harness\n    model: gpu-light\n    key_env: LITELLM_API_KEY",
+        ),
+    )
+    assert code == 1, out
+    assert "custom_providers[0].model = 'gpu-light'" in out
+    assert "RESULT: FAIL" in out
+
+
+def test_retired_raw_name_fails(tmp_path):
+    """Retired raw names no longer resolve (400), so they fail; the 2026-09-12 registry change moved qwen3.6-27B-code from raw-but-live to non-resolving."""
+    code, out = _run_config(
+        tmp_path,
+        "retired-qwen.yaml",
+        BASE.format(alias="gpu-vision").replace(
+            "delegation:\n  provider: harness",
+            "delegation:\n  provider: harness\n  model: qwen3.6-27B-code",
+        ),
+    )
+    assert code != 0, out
+    assert "delegation.model = 'qwen3.6-27B-code' is retired and no longer resolves" in out
+    assert "use gpu-dense" in out
+    assert "RESULT: FAIL" in out
+
+
+def test_retired_alias_in_fallback_providers_is_rejected(tmp_path):
+    """fallback_providers.model is model-bearing; a retired name there must fail."""
+    code, out = _run_config(
+        tmp_path,
+        "fallback-gpu-light.yaml",
+        BASE.format(alias="gpu-vision").replace("  model: deepseek-v4-flash", "  model: gpu-light"),
+    )
+    assert code == 1, out
+    assert "fallback_providers.model = 'gpu-light'" in out
+    assert "RESULT: FAIL" in out
+
+
+def test_retired_alias_in_x_search_is_rejected(tmp_path):
+    """x_search.model was previously not enumerated; the derivation must catch it."""
+    code, out = _run_config(
+        tmp_path,
+        "x-search-gpu-light.yaml",
+        BASE.format(alias="gpu-vision").replace(
+            "delegation:\n  provider: harness",
+            "delegation:\n  provider: harness\nx_search:\n  model: gpu-light",
+        ),
+    )
+    assert code == 1, out
+    assert "x_search.model = 'gpu-light'" in out
+    assert "RESULT: FAIL" in out
+
+
+def test_retired_alias_in_nested_auxiliary_block_is_rejected(tmp_path):
+    """A nested auxiliary sub-block outside the named three must still be derived."""
+    code, out = _run_config(
+        tmp_path,
+        "nested-aux-gpu-light.yaml",
+        BASE.format(alias="gpu-vision").replace(
+            "  compression:\n    model: syslog-auto\n    provider: harness\ndelegation:",
+            "  compression:\n    model: syslog-auto\n    provider: harness\n"
+            "  tasks:\n    summarize:\n      model: gpu-light\ndelegation:",
+        ),
+    )
+    assert code == 1, out
+    assert "auxiliary.tasks.summarize.model = 'gpu-light'" in out
+    assert "RESULT: FAIL" in out

tests/test_disk_gc_report_only.py

@@ -0,0 +1,185 @@
+"""Regression tests for the disk-gc report-only gate (CT 111 / tdunna / .129).
+
+WHY THIS FILE EXISTS: `disk-gc-threat-response.prose.md` defined AMBER as "GC scheduled
+for next run" and its Execution loop called `gc-executor` for EVERY threat, with no
+guest-level exclusion. CT 111 (tdunna, 192.168.68.129) belongs to Theo and is
+report-only per the captain (2026-08-17, re-confirmed 2026-09-10) — so a single AMBER
+reading on that guest would have scheduled GC commands (apt clean, journal vacuum,
+log/tmp deletion, snap removal) against someone else's box. The only marker was
+frontmatter `report_only_agents`, which names an AGENT while the scan unit is a GUEST.
+
+These tests execute the real planner (`scripts/disk-gc-plan.py`) and assert observable
+behaviour: an excluded guest never produces a `gc-executor` action at any level, while
+our own guests still do.
+"""
+from __future__ import annotations
+
+import json
+import pathlib
+import subprocess
+import sys
+
+ROOT = pathlib.Path(__file__).resolve().parent.parent
+PLAN = ROOT / "scripts" / "disk-gc-plan.py"
+
+
+def _plan(scan, tmp_path):
+    scan_file = tmp_path / "scan.json"
+    scan_file.write_text(json.dumps(scan))
+    proc = subprocess.run(
+        [sys.executable, str(PLAN), "--scan", str(scan_file), "--json"],
+        capture_output=True, text=True,
+    )
+    assert proc.returncode == 0, proc.stderr
+    return json.loads(proc.stdout)
+
+
+def _actions_for(plan, target):
+    return [row for row in plan if str(row["target"]) == str(target)]
+
+
+def test_excluded_guest_never_gets_gc_at_any_level(tmp_path):
+    """CT 111 at AMBER, RED and CRITICAL — always report-only, never gc-executor."""
+    for pct, level in ((84, "AMBER"), (90, "RED"), (97, "CRITICAL")):
+        plan = _plan([{"id": 111, "hostname": "tdunna", "ip": "192.168.68.129",
+                       "usage_pct": pct}], tmp_path)
+        rows = _actions_for(plan, 111)
+        assert rows, f"CT 111 must still be reported at {level}"
+        assert rows[0]["level"] == level
+        assert rows[0]["action"] == "report-only", rows
+        assert not any(r["action"] == "gc-executor" for r in rows)
+
+
+def test_exclusion_matches_on_any_identity_key(tmp_path):
+    """The gate is keyed on guest/host, so id, ct, ctid, hostname or IP all match."""
+    for entry in ({"id": 111, "usage_pct": 95},
+                  {"ct": 111, "usage_pct": 95},
+                  {"ctid": 111, "usage_pct": 95},
+                  {"hostname": "tdunna", "usage_pct": 95},
+                  {"ip": "192.168.68.129", "usage_pct": 95}):
+        plan = _plan([entry], tmp_path)
+        assert all(r["action"] == "report-only" for r in plan), (entry, plan)
+
+
+def test_exclusion_matches_encoded_identities(tmp_path):
+    """A differently-encoded CT 111 id must not slip past the gate to gc-executor."""
+    encodings = ({"id": 111.0},
+                 {"id": "111.0"},
+                 {"id": "lxc/111"},
+                 {"id": "qemu/111"},
+                 {"id": "0111"},
+                 {"id": "  111  "})
+    for alias in encodings:
+        plan = _plan([{**alias, "usage_pct": 97}], tmp_path)
+        assert plan, alias
+        assert plan[0]["action"] == "report-only", (alias, plan)
+
+
+def test_our_own_guests_still_get_gc(tmp_path):
+    """acerpve .9 and amdpve .15 are ours — they must still be acted on."""
+    plan = _plan([{"hostname": "acerpve", "ip": "192.168.68.9", "usage_pct": 77},
+                  {"hostname": "amdpve", "ip": "192.168.68.15", "usage_pct": 76}], tmp_path)
+    assert len(plan) == 2
+    assert all(r["action"] == "gc-executor" for r in plan), plan
+
+
+def test_below_threshold_emits_nothing(tmp_path):
+    """GREEN guests produce no action at all."""
+    assert _plan([{"id": 111, "usage_pct": 40}], tmp_path) == []
+
+
+def test_agent_name_alone_does_not_gate_a_guest(tmp_path):
+    """An agent-name marker must not be the gate: an unrelated guest still gets GC."""
+    plan = _plan([{"id": 999, "hostname": "koby", "usage_pct": 95}], tmp_path)
+    assert plan and plan[0]["action"] == "gc-executor"
+
+
+def test_unidentified_threat_fails_closed(tmp_path):
+    """A threshold-crossing entry with no recognized identity must not schedule GC."""
+    plan = _plan([{"usage_pct": 97}], tmp_path)
+    assert plan, "an unidentified threat must still be reported"
+    assert plan[0]["action"] == "report-only", plan
+    assert plan[0]["reason"], plan
+
+
+def test_exclusion_entry_without_identity_fails_closed(tmp_path):
+    """A mis-typed exclusion entry must break the run, never silently disable the gate."""
+    contract = tmp_path / "broken.prose.md"
+    contract.write_text(
+        "```yaml\n"
+        "report_only_guests:\n"
+        "  - node: storepve\n"
+        "    reason: \"typo - no guest identity\"\n"
+        "```\n"
+    )
+    scan_file = tmp_path / "scan.json"
+    scan_file.write_text(json.dumps([{"ct": 111, "usage_pct": 97}]))
+    proc = subprocess.run(
+        [sys.executable, str(PLAN), "--scan", str(scan_file),
+         "--contract", str(contract), "--json"],
+        capture_output=True, text=True,
+    )
+    assert proc.returncode != 0, proc.stdout
+    assert "gc-executor" not in proc.stdout
+    assert "identity" in proc.stderr.lower(), proc.stderr
+
+
+def test_empty_exclusion_block_fails_closed(tmp_path):
+    """An emptied report_only_guests list must break the run, not disable the gate."""
+    contract = tmp_path / "empty.prose.md"
+    contract.write_text("```yaml\nreport_only_guests: []\n```\n")
+    scan_file = tmp_path / "scan.json"
+    scan_file.write_text(json.dumps([{"ct": 111, "usage_pct": 97}]))
+    proc = subprocess.run(
+        [sys.executable, str(PLAN), "--scan", str(scan_file),
+         "--contract", str(contract), "--json"],
+        capture_output=True, text=True,
+    )
+    assert proc.returncode != 0, proc.stdout
+    assert "gc-executor" not in proc.stdout
+    assert "report-only gate" in proc.stderr.lower(), proc.stderr
+
+
+def _plan_with_contract(scan, contract_text, tmp_path, name):
+    contract = tmp_path / name
+    contract.write_text(contract_text)
+    scan_file = tmp_path / f"scan-{name}.json"
+    scan_file.write_text(json.dumps(scan))
+    proc = subprocess.run(
+        [sys.executable, str(PLAN), "--scan", str(scan_file),
+         "--contract", str(contract), "--json"],
+        capture_output=True, text=True,
+    )
+    assert proc.returncode == 0, proc.stderr
+    return json.loads(proc.stdout)
+
+
+def test_gate_is_read_from_the_contract_block(tmp_path):
+    """The gate is data-driven by the contract block: the planner excludes the guest
+    when the block names it and acts on it when the block does not. Executes the real
+    planner interface against both fixtures so the behaviour change is observable."""
+    scan = [{"id": 111, "hostname": "tdunna", "ip": "192.168.68.129", "usage_pct": 95}]
+    with_gate = (
+        "```yaml\n"
+        "report_only_guests:\n"
+        "  - guest: 111\n"
+        "    hostname: tdunna\n"
+        "    ip: 192.168.68.129\n"
+        "    reason: \"fixture reason\"\n"
+        "```\n"
+    )
+    without_gate = (
+        "```yaml\n"
+        "report_only_guests:\n"
+        "  - guest: 999\n"
+        "    reason: \"fixture excludes a different guest\"\n"
+        "```\n"
+    )
+
+    gated = _plan_with_contract(scan, with_gate, tmp_path, "gated.prose.md")
+    assert gated[0]["action"] == "report-only", gated
+    assert gated[0]["reason"] == "fixture reason", gated
+
+    ungated = _plan_with_contract(scan, without_gate, tmp_path, "ungated.prose.md")
+    assert ungated[0]["action"] == "gc-executor", ungated
+    assert ungated[0]["action"] != gated[0]["action"]